A method, device and equipment for constructing an opinion event subsiding prediction model

Abstract

The application provides a method, device and equipment for constructing an opinion event subsiding prediction model. The method comprises the following steps: obtaining a plurality of opinion network data of historical opinion events and preprocessing the opinion network data; obtaining a feature value of each feature corresponding to each preprocessed opinion network data based on a preset feature library, taking the feature value as a first feature set of the opinion network data; performing feature selection based on a Pearson correlation coefficient to obtain a second feature set, obtaining a third feature set of each opinion network data based on the second feature set of all opinion network data, and performing standardization processing to obtain a fourth feature set of each opinion network data; taking the fourth feature set of each opinion network data and a labeled true value as a data sample, training a PNN neural network model based on all obtained data samples, and completing construction of the opinion event subsiding prediction model. The prediction model obtained by the method can be used to more accurately predict whether an opinion event is in a subsiding period.

Description

Technical Field

[0001] This application relates to the field of Internet data processing technology, and in particular to a technology for constructing a predictive model for the fading of public opinion events. Background Technology

[0002] Public opinion events are typically hot-button issues or sudden social events that attract public attention and have a significant impact. With the rapid development of the internet, especially mobile internet, and the widespread use of smartphones, hot-button issues and sudden events in society attract widespread attention. People express and forward opinions online through text, images, videos, and even short links, and comment on others' expressed and forwarded opinions. This can evolve into public opinion events, generating a large amount of online data related to these events.

[0003] Generally, based on the development cycle of public opinion events, they can be roughly divided into the incubation period, formation period, fluctuation period, and decline period. The decline period signifies that people generally no longer pay attention to the event, and the corresponding public opinion ebbs, recedes, and solidifies. Numerous factors influence the formation, fluctuation, and decline of public opinion events, and these factors may interact and influence each other, making the process quite complex. Identifying the factors affecting public opinion events to predict their decline presents significant challenges, often requiring human intervention, such as manually determining the weight of each factor.

[0004] Therefore, how to predict whether a public opinion event is in a period of decline is the technical problem that this application aims to solve. Summary of the Invention

[0005] The purpose of this application is to provide a method, apparatus, and equipment for constructing a prediction model for the decline of public opinion events, which is used to predict whether a public opinion event is in the decline phase, so as to at least partially solve the technical problem in the prior art that it is difficult to predict whether a public opinion event is in the decline phase.

[0006] According to one aspect of this application, a method for constructing a predictive model for the fading of public opinion events is provided, wherein the method includes:

[0007] Based on a preset time threshold, several multi-source heterogeneous public opinion network data at different stages related to historical public opinion events that are already in the receding phase are acquired, and each multi-source heterogeneous public opinion network data is preprocessed to obtain multiple preprocessed multi-source heterogeneous public opinion network data.

[0008] Based on the preset feature library, feature quantization is performed on each preprocessed multi-source heterogeneous public opinion network data to obtain the feature value of each feature in the preset feature library corresponding to each preprocessed multi-source heterogeneous public opinion network data, which serves as the first feature set of each preprocessed multi-source heterogeneous public opinion network data.

[0009] Based on the Pearson correlation coefficient, feature selection is performed on the first feature set of each preprocessed multi-source heterogeneous public opinion network data to obtain the second feature set of the preprocessed multi-source heterogeneous public opinion network data. Based on the second feature sets of all preprocessed multi-source heterogeneous public opinion network data, a third feature set of each preprocessed multi-source heterogeneous public opinion network data is obtained. The feature values ​​of each feature in the third feature set of each preprocessed public opinion network data are standardized to obtain a fourth feature set of each preprocessed public opinion network data. The third feature set of each preprocessed public opinion network data includes several identical features and their feature values.

[0010] Based on the stage of the historical public opinion event corresponding to each preprocessed multi-source heterogeneous public opinion network data, the truth value is labeled for each preprocessed multi-source heterogeneous public opinion network data, and the fourth feature set and the corresponding truth value of each preprocessed multi-source heterogeneous public opinion network data are used as a data sample. The above operation is also performed on different historical public opinion events that are in the decline phase. Based on all the obtained data samples, a sample dataset is formed.

[0011] Based on the sample dataset, a PNN neural network model is trained. When the prediction accuracy of the trained PNN neural network model meets a preset threshold, the construction of the public opinion event fading prediction model is completed.

[0012] Optionally, the multi-source heterogeneous public opinion network data includes:

[0013] Data related to the aforementioned public opinion event from different information sources on the Internet, wherein the data includes textual and non-textual data.

[0014] Optionally, the plurality of features includes:

[0015] Features under several multi-dimensional dimensions, wherein the multi-dimensional dimensions include multiple first-level dimensions, each first-level dimension includes multiple second-level dimensions, and each second-level dimension includes several features.

[0016] Optionally, the plurality of first-level dimensions include:

[0017] Disseminator;

[0018] Content to be disseminated;

[0019] Motivation for dissemination;

[0020] Propagation and evolution.

[0021] Optionally, the second-level dimensions under the propagator include:

[0022] Media participation;

[0023] Influence of self-media;

[0024] Credibility of identity.

[0025] Optionally, several characteristics of the media engagement include: the amount of central-level media coverage;

[0026] Provincial media coverage;

[0027] The amount of media coverage at the prefecture-level city level;

[0028] Other media coverage.

[0029] Optionally, the features under the influence of the self-media include:

[0030] Percentage of original content created by personally verified users;

[0031] Percentage of reposts by personally verified users;

[0032] Percentage of comments from personally verified users;

[0033] Percentage of shares received by personally verified users;

[0034] Percentage of comments received by personally verified users;

[0035] Percentage of likes received by personally verified users.

[0036] Optionally, the features under the identity credibility include: the proportion of information about the location where the public opinion event occurred.

[0037] Optionally, the secondary dimensions under the dissemination content include:

[0038] Fineness;

[0039] Sentimental inclination;

[0040] Richness.

[0041] Optionally, the features at the level of refinement include:

[0042] Average word count per original post;

[0043] Average number of characters per forwarded message;

[0044] Average word count per comment.

[0045] Optionally, the features under the sentiment tendency include: positive original content and the number of reposts;

[0046] Neutral original content and repost volume;

[0047] Negative original content and reposts;

[0048] Number of positive comments;

[0049] Neutral comment volume;

[0050] Number of negative comments.

[0051] Optionally, the richness of features includes:

[0052] Only the proportion of information in short links is included;

[0053] This only includes the proportion of information in the video.

[0054] Only the information content of the images is included;

[0055] This includes the proportion of information in short links and videos;

[0056] The proportion of information including short links and images;

[0057] This includes the proportion of information presented in videos and images;

[0058] This includes the proportion of information in short links, videos, and images;

[0059] This includes the proportion of information in videos.

[0060] Optionally, the second-level dimensions under the propagation motivation include:

[0061] Motivation intensity;

[0062] Motivational expression level.

[0063] Optionally, several characteristics under the motivation intensity include: the proportion of extreme information.

[0064] Optionally, the features under the degree of motivational expression include:

[0065] Degree of behavioral intention expression;

[0066] Deep causal expression level;

[0067] The continued leaks and high level of attention;

[0068] Public attention to the outcome of the investigation.

[0069] Optionally, the secondary dimension under propagation and evolution includes: propagation degree;

[0070] Evolutionary degree.

[0071] Optionally, the features under the propagation degree include:

[0072] Original information content;

[0073] The amount of information forwarded;

[0074] Information content in the comments.

[0075] Optionally, the features under the degree of evolution include:

[0076] The rate of change in information volume compared to the previous period.

[0077] Optionally, the method for constructing a predictive model for the fading of public opinion events further includes:

[0078] Based on the preset time threshold, multi-source heterogeneous public opinion network data related to the public opinion event is acquired, and the multi-source heterogeneous public opinion network data is preprocessed to obtain preprocessed multi-source heterogeneous public opinion network data.

[0079] Based on the aforementioned features, feature quantization processing is performed on the preprocessed multi-source heterogeneous public opinion network data to obtain feature values ​​corresponding to the aforementioned features. Then, the feature values ​​of each feature are standardized to obtain the feature set of the preprocessed multi-source heterogeneous public opinion network data. The feature set includes the aforementioned features and their standardized feature values.

[0080] The feature set is input into the public opinion event decline prediction model, and based on the output of the public opinion event decline prediction model, it is determined whether the public opinion event is in the decline period.

[0081] According to another aspect of this application, an apparatus for constructing a predictive model for the fading of public opinion events is provided, wherein the apparatus comprises:

[0082] The first module is used to acquire several multi-source heterogeneous public opinion network data at different stages related to historical public opinion events that are already in the receding period, based on a preset time threshold, and to preprocess each multi-source heterogeneous public opinion network data to obtain multiple preprocessed multi-source heterogeneous public opinion network data.

[0083] The second module is used to perform feature quantization processing on each preprocessed multi-source heterogeneous public opinion network data based on a preset feature library, to obtain the feature value of each feature in the preset feature library corresponding to each preprocessed multi-source heterogeneous public opinion network data, which serves as the first feature set of each preprocessed multi-source heterogeneous public opinion network data.

[0084] The third module is used to perform feature selection on the first feature set of each preprocessed multi-source heterogeneous public opinion network data based on the Pearson correlation coefficient to obtain the second feature set of the preprocessed multi-source heterogeneous public opinion network data, and to obtain the third feature set of each preprocessed multi-source heterogeneous public opinion network data based on the second feature set of all preprocessed multi-source heterogeneous public opinion network data, and to standardize the feature values ​​of each feature in the third feature set of each preprocessed public opinion network data to obtain the fourth feature set of each preprocessed public opinion network data, wherein the third feature set of each preprocessed public opinion network data includes the same number of features and their feature values;

[0085] The fourth module is used to perform truth labeling on each preprocessed multi-source heterogeneous public opinion network data based on the stage of the historical public opinion event corresponding to each preprocessed multi-source heterogeneous public opinion network data, and to take the fourth feature set and the corresponding truth value of each preprocessed multi-source heterogeneous public opinion network data as a data sample. The above operation is also performed on different historical public opinion events that are in the decline phase. Based on all the obtained data samples, a sample dataset is formed.

[0086] The fifth module is used to train a PNN neural network model based on the sample dataset. When the prediction accuracy of the trained PNN neural network model meets the preset threshold, the construction of the public opinion event fading prediction model is completed.

[0087] Optionally, the apparatus for constructing a predictive model for the fading of public opinion events further includes:

[0088] The sixth module is used to acquire multi-source heterogeneous public opinion network data related to the public opinion event based on the preset time threshold, and to preprocess the multi-source heterogeneous public opinion network data to obtain preprocessed multi-source heterogeneous public opinion network data.

[0089] The seventh module is used to perform feature quantization processing on the preprocessed multi-source heterogeneous public opinion network data based on the aforementioned features, to obtain feature values ​​corresponding to the aforementioned features, and to standardize the feature values ​​of each feature to obtain the feature set of the preprocessed multi-source heterogeneous public opinion network data, wherein the feature set includes the aforementioned features and their standardized feature values.

[0090] The eighth module is used to input the feature set into the public opinion event decline prediction model, and to determine whether the public opinion event is in the decline period based on the output of the public opinion event decline prediction model.

[0091] Compared with existing technologies, this application provides a method, apparatus, and device for constructing a predictive model for the fading of public opinion events. The method includes: acquiring several multi-source heterogeneous public opinion network data at different stages related to historical public opinion events that are already in the fading phase, based on a preset time threshold; preprocessing each multi-source heterogeneous public opinion network data to obtain multiple preprocessed multi-source heterogeneous public opinion network data sets; performing feature quantization processing on each preprocessed multi-source heterogeneous public opinion network data based on a preset feature library to obtain the feature values ​​of each feature in the preset feature library corresponding to each preprocessed multi-source heterogeneous public opinion network data, serving as the first feature set of each preprocessed multi-source heterogeneous public opinion network data; performing feature selection on the first feature set of each preprocessed multi-source heterogeneous public opinion network data based on the Pearson correlation coefficient to obtain the second feature set of the preprocessed multi-source heterogeneous public opinion network data; and obtaining a model for each preprocessed multi-source heterogeneous public opinion network data based on the second feature sets of all preprocessed multi-source heterogeneous public opinion network data. The third feature set of the multi-source heterogeneous public opinion network data and the feature values ​​of each feature in the third feature set of each preprocessed public opinion network data are standardized to obtain the fourth feature set of each preprocessed public opinion network data. The third feature set of each preprocessed public opinion network data includes several identical features and their feature values. Based on the stage of the historical public opinion event corresponding to each preprocessed multi-source heterogeneous public opinion network data, the ground truth label is performed on each preprocessed multi-source heterogeneous public opinion network data. The fourth feature set and the corresponding ground truth of each preprocessed multi-source heterogeneous public opinion network data are used as a data sample. The above operation is performed on different historical public opinion events that are in the receding stage. Based on all the obtained data samples, a sample dataset is formed. Based on the sample dataset, a PNN (Probabilistic Neural Networks) neural network model is trained. When the prediction accuracy of the trained PNN neural network model meets the preset threshold, the construction of the public opinion event receding prediction model is completed. This method yields a model for predicting the decline of public opinion events. Without human intervention, it can objectively combine the acquired multi-source heterogeneous public opinion network data to accurately predict whether a public opinion event is in the decline phase. Attached Figure Description

[0092] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0093] Figure 1 A schematic flowchart of a method for constructing a predictive model for the fading of public opinion events, according to one aspect of this application, is shown.

[0094] Figure 2A schematic diagram of an apparatus for constructing a predictive model for the fading of public opinion events, according to another aspect of this application, is shown.

[0095] The same or similar reference numerals in the accompanying drawings represent the same or similar parts. Detailed Implementation

[0096] The present invention will now be described in further detail with reference to the accompanying drawings.

[0097] In a typical configuration of various embodiments of this application, each trusted party of the device, system and / or each module of the apparatus may include one or more processors (CPU), input / output interfaces, network interfaces and memory.

[0098] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0099] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include non-transitory computer-readable media, such as modulated data signals and carrier waves.

[0100] This application provides a technical solution for constructing a predictive model for the fading of public opinion events. By collecting public opinion network data from different sources on the Internet for a preset time period (e.g., 24 hours) at various stages (incubation period, formation period, fluctuation period, and decline period) of historical public opinion events that are already in the decline phase, several multi-source heterogeneous public opinion network data are obtained. Based on a preset feature library, feature quantification processing is performed on the multi-source heterogeneous public opinion network data to determine the feature value of each multi-source heterogeneous public opinion network data corresponding to the preset feature library. Feature selection is performed based on Pearson correlation coefficient, and further filtering is performed. The feature values ​​corresponding to the selected features are standardized to obtain the feature set of each multi-source heterogeneous public opinion network data. The feature set of each multi-source heterogeneous public opinion network data includes several identical features. The feature set of each multi-source heterogeneous public opinion network data and the labeled ground truth are used as a data sample to obtain data samples corresponding to different historical public opinion events that are already in the decline phase, resulting in a sample dataset. A PNN neural network model is trained to construct a public opinion event decline prediction model, which is used to predict whether public opinion events collected from the Internet are in the decline phase. The prediction process does not require manual intervention and has a high accuracy rate.

[0101] To further illustrate the technical means adopted and the effects achieved in this application, the technical solution of this application will be clearly and completely described below in conjunction with the accompanying drawings and preferred embodiments.

[0102] Figure 1 The diagram illustrates a method flow chart for constructing a predictive model for the fading of public opinion events according to one aspect of this application, wherein one embodiment of the method includes:

[0103] S101 acquires several multi-source heterogeneous public opinion network data at different stages related to historical public opinion events that are already in the decline phase, based on a preset time threshold, and preprocesses each multi-source heterogeneous public opinion network data to obtain multiple preprocessed multi-source heterogeneous public opinion network data.

[0104] S102 performs feature quantization on each preprocessed multi-source heterogeneous public opinion network data based on a preset feature library, and obtains the feature value of each feature in the preset feature library corresponding to each preprocessed multi-source heterogeneous public opinion network data, which serves as the first feature set of each preprocessed multi-source heterogeneous public opinion network data.

[0105] S103, based on the Pearson correlation coefficient, performs feature selection on the first feature set of each preprocessed multi-source heterogeneous public opinion network data to obtain the second feature set of the preprocessed multi-source heterogeneous public opinion network data. Based on the second feature sets of all preprocessed multi-source heterogeneous public opinion network data, a third feature set of each preprocessed multi-source heterogeneous public opinion network data is obtained. Furthermore, the feature values ​​of each feature in the third feature set of each preprocessed public opinion network data are standardized to obtain a fourth feature set of each preprocessed public opinion network data. The third feature set of each preprocessed public opinion network data includes several identical features and their feature values.

[0106] S104, based on the stage of the historical public opinion event corresponding to each preprocessed multi-source heterogeneous public opinion network data, performs truth value labeling on each preprocessed multi-source heterogeneous public opinion network data, and takes the fourth feature set and corresponding truth value of each preprocessed multi-source heterogeneous public opinion network data as a data sample. The above operation is also performed on different historical public opinion events that are in the decline phase, and a sample dataset is formed based on all the obtained data samples.

[0107] S105 Based on the sample dataset, train a PNN neural network model. When the prediction accuracy of the trained PNN neural network model meets the preset threshold, the construction of the public opinion event fading prediction model is completed.

[0108] The various method embodiments / optional embodiments of this application can be implemented or executed by device 100, wherein device 100 is a computer device with corresponding hardware and software environment. The computer device includes, but is not limited to, personal computers, laptops, industrial computers, servers, network hosts, single network servers, or network server clusters. Here, the computer device is merely an example; other existing or future devices and / or resource platforms that are applicable to this application should also be included within the scope of protection of this application, and are hereby incorporated by reference.

[0109] In this embodiment, in step S101, the device 100 can, for several historical public opinion events that are already in the decline phase, acquire several multi-source heterogeneous public opinion network data related to each stage of the historical public opinion events that are already in the decline phase, based on one or more keywords associated with each historical event and according to a preset time threshold. For example, according to the development timeline of the historical public opinion event, starting from the incubation period, covering the formation period, fluctuation period, and decline period, the preset time threshold is set to 24 hours, and data is collected sequentially. The public opinion network data collected from multiple information sources on the Internet is divided into 24-hour units, and the public opinion network data from multiple information sources within each 24-hour period is taken as a multi-source heterogeneous public opinion network data. Each multi-source heterogeneous public opinion network data is then preprocessed to obtain preprocessed multi-source heterogeneous public opinion network data.

[0110] The preprocessing of the acquired multi-source heterogeneous public opinion network data includes data cleaning and data integration. Data that is irrelevant to the public opinion event and invalid data are removed from the acquired multi-source heterogeneous public opinion network data. Then, the cleaned data is integrated and summarized.

[0111] Optionally, the multi-source heterogeneous public opinion network data includes:

[0112] Data related to the aforementioned public opinion event from different information sources on the Internet, wherein the data includes textual and non-textual data.

[0113] The device 100 can collect network information containing relevant keywords or combinations of relevant keywords from different internet sources according to a preset time threshold, based on one or more keywords associated with the public opinion event, to obtain data related to the public opinion event, serving as multi-source heterogeneous public opinion network data. These different internet sources can include clients, websites, videos, electronic newspapers and / or social networking platforms such as Weibo and WeChat official accounts. The multi-source public opinion network data collected from different internet sources includes plain text data, such as user-generated or forwarded text comments, as well as non-text data, such as user-posted images, short links, and videos.

[0114] Continuing in this embodiment, in step S102, the device 100 can perform feature quantization processing on each preprocessed multi-source heterogeneous public opinion network data based on a preset feature library, and obtain the feature value of each feature in the preset feature library corresponding to each preprocessed multi-source heterogeneous public opinion network data, as the first feature set of each preprocessed multi-source heterogeneous public opinion network data.

[0115] This involves using professional knowledge and practical experience, combined with the analysis of different historical public opinion events, to pre-summarize a feature library influencing the development of public opinion events. For each feature in this feature library, feature quantification is performed on each pre-processed multi-source heterogeneous public opinion network data set to obtain the feature value of each feature in the pre-defined feature library for each pre-processed multi-source heterogeneous public opinion network data set. This value can serve as the first feature set for each pre-processed multi-source heterogeneous public opinion network data set. The quantification method used is related to the data type of the public opinion network data set and can be statistical values, average values, or percentages, etc., and is not limited here.

[0116] Continuing in this embodiment, in step S103, device 100 may perform feature selection on the first feature set of each preprocessed multi-source heterogeneous public opinion network data based on the Pearson correlation coefficient to obtain the second feature set of the preprocessed multi-source heterogeneous public opinion network data, and obtain the third feature set of each preprocessed multi-source heterogeneous public opinion network data based on the second feature set of all preprocessed multi-source heterogeneous public opinion network data, and perform standardization processing on the feature values ​​of each feature in the third feature set of each preprocessed public opinion network data to obtain the fourth feature set of each preprocessed public opinion network data, wherein the third feature set of each preprocessed public opinion network data includes the same number of features and their feature values.

[0117] First, the Pearson correlation coefficient between each feature of the first feature set of each preprocessed multi-source heterogeneous public opinion network data and the target variable (or truth variable, i.e., the variable indicating whether the public opinion event has subsided; when the public opinion event corresponding to the preprocessed multi-source heterogeneous public opinion network data is in the subsidence phase, the target variable takes a value of 1, otherwise it takes a value of 0) is calculated. The Pearson correlation coefficient of each feature is compared with a first preset threshold. If it is greater than the preset threshold, it is retained; otherwise, it is deleted. Then, the Pearson correlation coefficient between each pair of features retained in the first feature set of each preprocessed multi-source heterogeneous public opinion network data is calculated, and compared with a second preset threshold. If it is greater than the preset threshold, it indicates that the two features are highly correlated, and only the feature with the larger Pearson correlation coefficient with the target variable needs to be retained; the other feature can be deleted. After iterating through the first feature set of each preprocessed multi-source heterogeneous public opinion network data, features highly correlated with the development of public opinion events are selected, and redundant features are removed. The retained features and their feature values ​​are used as the second feature set of each preprocessed multi-source heterogeneous public opinion network data. The second feature set may include different features for each preprocessed multi-source heterogeneous public opinion network data. The Pearson correlation coefficient is used to measure the linear correlation between two features, with a value range of [-1, 1], where 1 represents a perfect positive correlation, -1 represents a perfect negative correlation, and 0 represents no correlation.

[0118] Then, for the second feature set of all preprocessed multi-source heterogeneous public opinion network data, a chi-square test is used for feature selection. Specifically, the chi-square test is used to calculate the chi-square value between each feature in the second feature set of each preprocessed multi-source heterogeneous public opinion network data and the target variable. Features whose chi-square values ​​satisfy preset conditions are retained. This process is repeated for each feature in the second feature set of each preprocessed multi-source heterogeneous public opinion network data to obtain the third feature set of each preprocessed multi-source heterogeneous public opinion network data. The features in the third feature set of each preprocessed public opinion network data are the same, including the aforementioned features across multiple dimensions, but the feature values ​​of each feature may differ (features with a value of zero may exist).

[0119] Finally, the feature values ​​of each feature in the third feature set of each preprocessed multi-source heterogeneous public opinion network data may have different dimensions and different data value ranges. In order to eliminate the influence of dimensions and data value ranges, reduce the complexity of data processing, and retain the relationships existing in the original data, the feature values ​​of each feature in the third feature set of each preprocessed multi-source heterogeneous public opinion network data can be standardized. The resulting fourth feature set includes several features and their standardized feature values. The feature value of each feature in the fourth feature set is in the interval [0,1].

[0120] Optionally, the plurality of features includes:

[0121] Features under several multi-dimensional dimensions, wherein the multi-dimensional dimensions include multiple first-level dimensions, each first-level dimension includes multiple second-level dimensions, and each second-level dimension includes several features.

[0122] Among them, based on expert knowledge, experience, and the characteristics of multi-source heterogeneous public opinion network data, combined with relevant thresholds, several features with high correlation to the development stage of public opinion events, especially the decline stage, can be determined. These features include features under several multi-dimensional dimensions, where each multi-dimensional dimension includes multiple first-level dimensions, each first-level dimension includes multiple second-level dimensions, and each second-level dimension includes several features.

[0123] Optionally, the plurality of first-level dimensions include:

[0124] Disseminator;

[0125] Content to be disseminated;

[0126] Motivation for dissemination;

[0127] Propagation and evolution.

[0128] Based on expert knowledge, experience, and the characteristics of multi-source heterogeneous public opinion network data, first-level dimensions such as communicators, content, motivation, and evolution of the public opinion network data can be set.

[0129] Optionally, the second-level dimensions under the propagator include:

[0130] Media participation;

[0131] Influence of self-media;

[0132] Credibility of identity.

[0133] Within the primary dimension of communicators, secondary dimensions are set up, including media participation, self-media influence, and identity feasibility.

[0134] Optionally, the features of media engagement include:

[0135] The amount of media coverage at the central level;

[0136] Provincial media coverage;

[0137] The amount of media coverage at the prefecture-level city level;

[0138] Other media coverage.

[0139] Based on the secondary dimension of media participation, the pre-processed multi-source heterogeneous public opinion network data is classified according to its source. The public opinion network data from the media can be divided into four categories according to the media level: central media, provincial media, municipal media, and other media. Then, by counting the media coverage of central media, provincial media, municipal media, and other media, four characteristics with high correlation to the decline of public opinion events under the secondary dimension of media participation can be identified: the media coverage of central media, provincial media, municipal media, and other media.

[0140] Optionally, the features under the influence of the self-media include:

[0141] Percentage of original content created by personally verified users;

[0142] Percentage of reposts by personally verified users;

[0143] Percentage of comments from personally verified users;

[0144] Percentage of shares received by personally verified users;

[0145] Percentage of comments received by personally verified users;

[0146] Percentage of likes received by personally verified users.

[0147] Based on the secondary dimension of self-media influence, the data is categorized according to the source of the pre-processed multi-source heterogeneous public opinion network data. Data originating from verified accounts is further categorized by "account verification type." The proportions of original content, reposts, and comments from verified individual users are then statistically analyzed (i.e., the proportions of original content, reposts, and comments from verified individual users within the public opinion network data originating from verified accounts) to assess the role of high-influence users. Furthermore, the proportions of reposts (including reposts of original content, and potentially reposts of reposts), comments, and likes of verified individual user public opinion network data are statistically analyzed to reflect the influence of high-influence users on the development of online public opinion events. This allows for the identification of six characteristics highly correlated with the decline of public opinion events under the secondary dimension of self-media influence: the proportions of original content, reposts, comments, reposts, comments, and likes from verified individual users.

[0148] Optionally, the features under the identity credibility include:

[0149] The proportion of information about the location where the public opinion event occurred.

[0150] Among them, based on the secondary dimension of identity credibility, the geographical information of the publisher of each piece of public opinion network data can be determined according to the network information contained in the preprocessed multi-source heterogeneous public opinion network data. The number of public opinion network data whose geographical information is the location of the public opinion event (the location of the public opinion event can be determined in advance as a parameter setting, or existing technologies can be used, such as the method disclosed in "A Method and Device for Geographical Identification of Public Opinion Information", Publication No. CN 103064951B) can be used to determine the characteristic of the proportion of public opinion event location information that is highly relevant to the decline period of the public opinion event under the secondary dimension of identity credibility.

[0151] Optionally, the secondary dimensions under the dissemination content include:

[0152] Fineness;

[0153] Sentimental inclination;

[0154] Richness.

[0155] Within the primary dimension of content dissemination, secondary dimensions are set up, including refinement, emotional inclination, and richness.

[0156] Optionally, the features at the level of refinement include:

[0157] Average word count per original post;

[0158] Average number of characters per forwarded message;

[0159] Average word count per comment.

[0160] Based on the secondary dimension of granularity, the preprocessed multi-source heterogeneous public opinion network data is classified into types such as original, forwarded, and comment. Then, the number of words in the text content of each piece of public opinion network data is counted (for non-text type public opinion network data, such as images, videos, and short links, the text content can be identified first, and then the word count can be counted). This allows us to determine three features that are highly correlated with the decline period of public opinion events under the secondary dimension of granularity: the average number of words in each original message, the average number of words in each forwarded message, and the average number of words in each comment message.

[0161] Optionally, the features under the sentiment tendency include:

[0162] Positive original content and reposts;

[0163] Neutral original content and repost volume;

[0164] Negative original content and reposts;

[0165] Number of positive comments;

[0166] Neutral comment volume;

[0167] Number of negative comments.

[0168] Based on the secondary dimension of sentiment tendency, on the one hand, each piece of public opinion network data can be input into a pre-trained public opinion information sentiment tendency prediction model in text form (for non-text public opinion network data, the text must be identified first) to determine the sentiment tendency attribute of the public opinion network data, including positive, neutral, and negative. On the other hand, depending on whether the public opinion network data is original, forwarded, or commented, six features with high correlation to the decline of public opinion events under the secondary dimension of sentiment tendency can be obtained: positive original and forwarded volume, neutral original and forwarded volume, negative original and forwarded volume, positive comment volume, neutral comment volume, and negative comment volume.

[0169] Optionally, the richness of features includes:

[0170] Only the proportion of information in short links is included;

[0171] This only includes the proportion of information in the video.

[0172] Only the information content of the images is included;

[0173] This includes the proportion of information in short links and videos;

[0174] The proportion of information including short links and images;

[0175] This includes the proportion of information presented in videos and images;

[0176] This includes the proportion of information in short links, videos, and images;

[0177] This includes the proportion of information in videos.

[0178] Based on the richness dimension, the acquired public opinion network data containing non-textual data was classified, and the number of information items including only short links, only videos, only images, short links and videos, short links and images, videos and images, short links, videos and images, and videos were counted respectively. This yielded eight characteristics under the richness dimension that were highly relevant to the decline of public opinion events: the proportion of information including only short links, the proportion of information including only videos, the proportion of information including only images, the proportion of information including short links and videos, the proportion of information including short links and images, the proportion of information including videos and images, the proportion of information including short links, videos and images, and the proportion of information including videos.

[0179] Optionally, the second-level dimensions under the propagation motivation include:

[0180] Motivation intensity;

[0181] Motivational expression level.

[0182] Within the primary dimension of communication motivation, secondary dimensions such as motivation intensity and motivation expression are set.

[0183] Optionally, the features under the motivation intensity include:

[0184] The proportion of extreme information.

[0185] Based on the secondary dimension of motivation intensity, extreme word identification can be performed on each piece of public opinion network data. This involves pre-constructing a dictionary containing several extreme words, traversing the text content of each piece of public opinion network data (for non-text public opinion network data, the text itself can be identified first), and determining whether it contains at least one extreme word from the dictionary. If it does, the piece of public opinion network data can be considered extreme information. After traversing each piece of public opinion network data, the number of data points containing extreme words is counted, yielding the characteristic of the proportion of extreme information highly correlated with the decline period of public opinion events under the secondary dimension of motivation intensity.

[0186] Optionally, the features under the degree of motivational expression include:

[0187] Degree of behavioral intention expression;

[0188] Deep causal expression level;

[0189] The continued leaks and high level of attention;

[0190] Public attention to the outcome of the investigation.

[0191] Based on the secondary dimension of motivational expression, keyword identification can be performed on each piece of public opinion network data. Using relevant sociological and communication theories, keywords associated with four motivational aspects—expression of behavioral intention, in-depth causal investigation, continuous revelations, and handling results—can be pre-defined. These keywords can be expanded using a text analysis clustering model and then filtered by experts to obtain dictionaries corresponding to each of the four motivations. The text content of each piece of public opinion network data (for non-textual data, the text itself can be identified first) is traversed to determine if it contains at least one keyword from the dictionary. If it does, the piece of public opinion network data is considered associated with the corresponding motivation. After traversing each piece of public opinion network data, the number of data points associated with each motivation is counted, yielding four characteristics highly correlated with the decline of public opinion events under the secondary dimension of motivational expression: expression of behavioral intention, expression of in-depth causal investigation, attention to continuous revelations, and attention to handling results.

[0192] Optionally, the second-level dimensions under propagation and evolution include:

[0193] Dissemination;

[0194] Evolutionary degree.

[0195] Within the primary dimension of propagation and evolution, secondary dimensions such as propagation degree and evolution degree are set.

[0196] Optionally, the features under the propagation degree include:

[0197] Original information content;

[0198] The amount of information forwarded;

[0199] Information content in the comments.

[0200] Based on the secondary dimension of dissemination, the preprocessed multi-source heterogeneous public opinion network data is classified into original, forwarded, and commented information, and the amount of original information, forwarded information, and commented information is statistically analyzed. This allows us to determine three characteristics that are highly correlated with the decline of public opinion events under the secondary dimension of dissemination: the amount of original information, forwarded information, and commented information.

[0201] Optionally, the features under the degree of evolution include:

[0202] The rate of change in information volume compared to the previous period.

[0203] Among them, based on the second-level dimension of evolution degree, the ratio of the difference between the number of pre-processed multi-source heterogeneous public opinion network data in the current period and the number of pre-processed multi-source heterogeneous public opinion network data in the previous period to the number of pre-processed multi-source heterogeneous public opinion network data in the previous period can determine the feature of the month-on-month change rate of information volume that is highly correlated with the decline period of public opinion events under the second-level dimension of evolution degree.

[0204] In step S103, the third feature set of each preprocessed public opinion network data in an optional embodiment includes the same 37 features as shown in Table 1 below.

[0205] Table 1

[0206]

[0207]

[0208] For example, by combining preset thresholds, several features can be determined, as shown in Table 1, which contains 37 features across three dimensions. That is, the third feature set of each preprocessed public opinion network data set includes the same 37 features, but their corresponding 37 feature values ​​may differ. After standardization, a fourth feature set can be obtained for each preprocessed public opinion network data set that is highly correlated with the decline period of the public opinion event. This fourth feature set includes the 37 features shown in Table 1 and their corresponding 37 standardized feature values.

[0209] The standardization of the feature values ​​of each feature in the third feature set of each preprocessed multi-source heterogeneous public opinion network data can be accomplished by combining the sample dataset. For example, if the third feature set of each preprocessed public opinion network data includes the 37 features shown in Table 1, the feature values ​​of each feature in the third feature set of each preprocessed multi-source heterogeneous public opinion network data can be standardized using the Min-Max Normalization method, combined with the sample dataset. The process is as follows: Let y mi Let x be the unstandardized feature value of the m-th feature in the i-th data sample. mi The corresponding standardized eigenvalues, x mj Let m be the feature value of the m-th feature in the j-th data sample before standardization. Standardization can be performed according to the following formula (1):

[0210]

[0211] Where n is the number of data samples in the sample dataset, min[x mj} represents the minimum value among all unstandardized feature values ​​of the m-th feature in the sample dataset, max{x} mj} represents the maximum value among all unstandardized feature values ​​of the m-th feature in the sample dataset.

[0212] Continuing in this embodiment, in step S104, the device 100 can perform truth labeling on each preprocessed multi-source heterogeneous public opinion network data based on the stage of the historical public opinion event corresponding to each preprocessed multi-source heterogeneous public opinion network data, and take the fourth feature set and the corresponding truth value of each preprocessed multi-source heterogeneous public opinion network data as a data sample, and perform the above operations on different historical public opinion events that are in the decline period, and form a sample dataset based on all the obtained data samples.

[0213] Specifically, based on the stage of the historical public opinion event corresponding to each multi-source heterogeneous public opinion network data, a truth value label is made for each multi-source heterogeneous public opinion network data, and the fourth feature set and the labeled truth value of each multi-source heterogeneous public opinion network data are used as a data sample.

[0214] Based on professional knowledge and experience, the development timeline of historical public opinion events can be analyzed to determine the stage (latent period, formation period, fluctuation period, and decline period) of the historical public opinion event corresponding to each multi-source heterogeneous public opinion network data. Therefore, according to the stage of the historical public opinion event corresponding to each multi-source heterogeneous public opinion network data, a truth value can be marked for each multi-source heterogeneous public opinion network data. For example, if the stage of the historical public opinion event corresponding to the multi-source heterogeneous public opinion network data is the decline period, then the truth value is marked as 1; otherwise, the truth value is marked as 0.

[0215] In order to obtain a diverse range of data samples that can meet the needs of building a predictive model for the decline of public opinion events, the above operations can be performed on several different historical public opinion events that are already in the decline phase to obtain a sufficient number of data samples to meet the requirements, and all the obtained data samples can be combined into a sample dataset.

[0216] It should be noted that, based on a pre-set feature library, relevant thresholds can be set for different stages of public opinion events, and annotations can be performed to construct prediction models for different stages of different public opinion events. Specifically, several features for different stages can be determined by pre-setting different thresholds, thus determining the fourth feature set for each pre-processed multi-source heterogeneous public opinion network data. The fourth feature set and the labeled true values ​​of each pre-processed multi-source heterogeneous public opinion network data are used as corresponding data samples to train the model predicting the corresponding stage. For example, for a fluctuating period, after determining the stage of the historical public opinion event corresponding to each multi-source heterogeneous public opinion network data, the fourth feature set of each multi-source heterogeneous public opinion network data is determined by pre-setting different thresholds. If the stage of the historical public opinion event corresponding to the multi-source heterogeneous public opinion network data is a fluctuating period, the labeled true value is 1; otherwise, the labeled true value is 0. This data sample can then be used to construct a prediction model for the fluctuating period of public opinion events.

[0217] Continuing in this embodiment, in step S105, the device 100 can train a PNN neural network model based on the sample dataset. When the prediction accuracy of the trained PNN neural network model meets a preset threshold, the construction of the public opinion event fading prediction model is completed.

[0218] Based on the obtained sample dataset, a PNN neural network model is trained. When the prediction accuracy of the trained PNN neural network model meets the preset threshold, the construction of the public opinion event fading prediction model is completed.

[0219] The obtained sample dataset can be divided into a training set, a validation set, and a test set. For example, data samples in the sample dataset can be randomly selected as training, validation, or test samples in an 8:1:1 ratio. First, the training set is used, and the data samples in the training set are input into a PNN neural network model with relevant training parameters for training. Several PNN neural network models can be obtained after training, depending on the set training parameters. Next, the validation set is used for validation, and the PNN neural network model with the best validation performance is used as the trained PNN neural network model. Then, the test set is used, and the feature values ​​of the data samples in the test set are input into the trained PNN neural network model. The predicted value output by the model is compared with the true value of the data sample. If they match, the prediction is accurate. This process is repeated for each data sample in the test set. When the prediction accuracy based on the test set meets a preset threshold, the construction of the public opinion event fading prediction model is complete, and this trained PNN neural network model can be used as the public opinion event fading prediction model. If the test accuracy based on the test set does not meet the preset threshold, the PNN neural network model can be retrained based on the training set and validation set according to the relevant training parameters. After the retrained PNN neural network model is determined, it can be tested again based on the test set until the prediction accuracy of the retrained PNN neural network model meets the preset threshold.

[0220] Among them, the PNN neural network model possesses the ability to classify arbitrarily complex patterns and excellent multidimensional function mapping capabilities, enabling it to solve XOR and other problems that simple perceptrons cannot handle. Structurally, the PNN neural network model has an input layer, hidden layers, and an output layer. Essentially, the PNN neural network model is an algorithm that uses the squared error as the objective function and employs gradient descent to calculate the minimum value of the objective function.

[0221] An example is a PNN neural network model built using the open-source machine learning library PyTorch, comprising an input layer, three hidden layers, and an output layer. Based on 1017 preprocessed multi-source heterogeneous public opinion network data points from 77 historical public opinion events that are currently in their decline phase, collected and preprocessed over 24-hour periods, feature quantization is performed on each preprocessed multi-source heterogeneous public opinion network data point using a pre-defined feature library. This yields the feature values ​​of each feature in the pre-defined feature library for each preprocessed multi-source heterogeneous public opinion network data point, serving as the first feature set for each preprocessed multi-source heterogeneous public opinion network data point. Then, based on the Pearson correlation coefficient, feature selection is performed on the first feature set of each preprocessed multi-source heterogeneous public opinion network data point to obtain the second feature set of the preprocessed multi-source heterogeneous public opinion network data point. Based on the second feature set of all preprocessed multi-source heterogeneous public opinion network data and a preset threshold, a third feature set is obtained for each preprocessed multi-source heterogeneous public opinion network data. The features in the third feature set of each preprocessed public opinion network data are the same, including the aforementioned features, but the feature values ​​of each feature may be different. The feature values ​​of each feature in the third feature set of each preprocessed public opinion network data are standardized to obtain a fourth feature set for each preprocessed public opinion network data. Based on the stage of the historical public opinion event corresponding to each preprocessed multi-source heterogeneous public opinion network data, ground truth labeling is performed on each preprocessed multi-source heterogeneous public opinion network data. For example, the fourth feature set of historical public opinion network data with high correlation to the decline period of public opinion events can be obtained, including 37 features as shown in Table 1 and their corresponding 37 standardized feature values. Combined with the stage of the historical public opinion event corresponding to each multi-source heterogeneous public opinion network data, the ground truth label is performed on each multi-source heterogeneous public opinion network data (the stage of the corresponding historical public opinion event is the decline period, and the ground truth value is 1; otherwise, the ground truth value is 0). The fourth feature set of each preprocessed multi-source heterogeneous public opinion network data, including the standardized feature values ​​of these 37 features and the corresponding ground truth value, is used as a data sample to form a sample dataset including 1017 data samples. After training, testing, and verification of the PNN neural network model with the above structure and parameter values ​​set in Table 2, a trained PNN neural network model with a prediction accuracy of over 85% can be obtained. This trained PNN neural network model is used as a public opinion event decline prediction model to predict whether public opinion events on the Internet will decline.

[0222] Table 2

[0223] Number of iterations 1000 Learning rate 0.0009 loss function CrossEntropyLoss Optimizer AdamW

[0224] Optionally, the aforementioned public opinion event fading prediction model can also be used to predict whether a public opinion event is in a fading phase, wherein, for example... Figure 1 One alternative embodiment of the method for constructing a predictive model for the fading of public opinion events, as shown, further includes:

[0225] S106 Based on the preset time threshold, acquire multi-source heterogeneous public opinion network data related to the public opinion event, and preprocess the multi-source heterogeneous public opinion network data to obtain preprocessed multi-source heterogeneous public opinion network data.

[0226] S107 Based on the aforementioned features, feature quantization processing is performed on the preprocessed multi-source heterogeneous public opinion network data to obtain feature values ​​corresponding to the aforementioned features, and each feature value is standardized to obtain a feature set of the preprocessed multi-source heterogeneous public opinion network data, wherein the feature set includes the aforementioned features and their standardized feature values.

[0227] S108 Input the feature set into the public opinion event decline prediction model, and determine whether the public opinion event is in the decline period based on the output of the public opinion event decline prediction model.

[0228] In this optional embodiment, in step S106, the device 100 can collect network information containing keywords or combinations of keywords related to the public opinion event based on one or more keywords associated with the public opinion event, using the same preset time threshold as when constructing the public opinion event decline prediction model, to obtain multi-source heterogeneous public opinion network data related to the public opinion event in units of the preset time threshold. For example, if the preset time threshold is 24 hours, multi-source heterogeneous public opinion network data related to the public opinion event can be obtained within 24 hours before the current time point or the time point to be monitored, based on one or more keywords related to the public opinion event, and the obtained multi-source heterogeneous public opinion network data can be preprocessed to obtain preprocessed multi-source heterogeneous public opinion network data.

[0229] The preprocessing of the acquired multi-source heterogeneous public opinion network data includes data cleaning and data integration. Data that is irrelevant to the public opinion event and invalid data are removed from the acquired multi-source heterogeneous public opinion network data. Then, the cleaned data is integrated and summarized.

[0230] Continuing in this optional embodiment, in step S107, the device 100 may perform feature quantization processing on the preprocessed multi-source heterogeneous public opinion network data based on the plurality of features to obtain feature values ​​corresponding to the plurality of features, and perform standardization processing on each feature value to obtain a feature set of the preprocessed multi-source heterogeneous public opinion network data, wherein the feature set includes the plurality of features and their standardized feature values.

[0231] In this process, the same features used in the training of the public opinion event fading prediction model constructed in the aforementioned embodiments can be used to perform feature quantization on the preprocessed multi-source heterogeneous public opinion network data. This yields the feature value of each feature corresponding to the preprocessed multi-source heterogeneous public opinion network data. The quantization method is related to the data type of the public opinion network and can be a statistical value, average value, or percentage, etc. Since the feature values ​​of features under different dimensions may have different dimensions and the data value ranges may vary significantly, in order to eliminate the influence of dimensions and data value ranges and retain the relationships existing in the original data, the feature values ​​of each feature of the preprocessed multi-source heterogeneous public opinion network data can be standardized using the sample dataset used in constructing the public opinion event fading prediction model. This yields a feature set of the preprocessed multi-source heterogeneous public opinion network data, which includes the aforementioned features and their standardized feature values. Each standardized feature value is within the range [0,1]. For example, if the features used in the training process of the public opinion event fading prediction model constructed in the aforementioned embodiments are the 37 features shown in Table 1, after performing feature quantization on the preprocessed multi-source heterogeneous public opinion network data, the corresponding 37 feature values ​​are obtained. Combined with the sample dataset used in constructing the public opinion event fading prediction model, each feature value is standardized to obtain a feature set of the preprocessed multi-source heterogeneous public opinion network data, including the 37 features and their standardized feature values. For example, the Min-Max Normalization method can be used to standardize each feature value using the sample dataset used in constructing the public opinion event fading prediction model. Let y m The unstandardized eigenvalue x of the m-th feature m The corresponding standardized eigenvalues, x mj The unstandardized feature value of the m-th feature in the j-th data sample of the pre-constructed public opinion event fading prediction model is used. Standardization can be performed according to the following formula (2):

[0232]

[0233] Where n is the number of data samples in the sample dataset, min{x mj} represents the minimum value among all unstandardized feature values ​​of the m-th feature in the sample dataset, max{x} mj} represents the maximum value among all unstandardized feature values ​​of the m-th feature in the sample dataset.

[0234] In the above formula (2), when determining the maximum and minimum values ​​of the feature values ​​before standardization, the sample dataset used can be a sample dataset that includes all samples, or a training set that only includes sample data used for training.

[0235] Continuing in this embodiment, in step S108, the device 100 can input the several standardized feature values ​​into a pre-built public opinion event decline prediction model, and determine whether the public opinion event is in the decline period based on the output of the public opinion event decline prediction model.

[0236] Specifically, the feature set is input into the public opinion event decline prediction model to predict whether the public opinion event corresponding to the preprocessed multi-source heterogeneous public opinion network data is in the decline period. Based on the output of the public opinion event decline model, it can be determined whether the public opinion event is in the decline period.

[0237] Figure 2 The diagram illustrates an apparatus for constructing a predictive model for the fading of public opinion events according to another aspect of this application, wherein, in one embodiment, the apparatus includes:

[0238] The first module 210 is used to acquire several multi-source heterogeneous public opinion network data at different stages related to historical public opinion events that are already in the decline phase, based on a preset time threshold, and to preprocess each multi-source heterogeneous public opinion network data to obtain multiple preprocessed multi-source heterogeneous public opinion network data.

[0239] The second module 220 is used to perform feature quantization processing on each preprocessed multi-source heterogeneous public opinion network data based on a preset feature library, to obtain the feature value of each feature in the preset feature library corresponding to each preprocessed multi-source heterogeneous public opinion network data, which serves as the first feature set of each preprocessed multi-source heterogeneous public opinion network data.

[0240] The third module 230 is used to perform feature selection on the first feature set of each preprocessed multi-source heterogeneous public opinion network data based on the Pearson correlation coefficient to obtain the second feature set of the preprocessed multi-source heterogeneous public opinion network data, and to obtain the third feature set of each preprocessed multi-source heterogeneous public opinion network data based on the second feature set of all preprocessed multi-source heterogeneous public opinion network data, and to standardize the feature values ​​of each feature in the third feature set of each preprocessed public opinion network data to obtain the fourth feature set of each preprocessed public opinion network data, wherein the third feature set of each preprocessed public opinion network data includes the same number of features and their feature values;

[0241] The fourth module 240 is used to perform truth labeling on each preprocessed multi-source heterogeneous public opinion network data based on the stage of the historical public opinion event corresponding to each preprocessed multi-source heterogeneous public opinion network data, and to take the fourth feature set and the corresponding truth value of each preprocessed multi-source heterogeneous public opinion network data as a data sample. The above operation is also performed on different historical public opinion events that are in the decline phase. Based on all the obtained data samples, a sample dataset is formed.

[0242] The fifth module 250 is used to train a PNN neural network model based on the sample dataset. When the prediction accuracy of the trained PNN neural network model meets the preset threshold, the construction of the public opinion event receding prediction model is completed.

[0243] In this embodiment, the device is deployed or integrated in the device 100 that performs the foregoing method embodiments and / or optional embodiments.

[0244] In this embodiment, the first module 210 of the device can, for several historical public opinion events that are already in the decline phase, acquire several multi-source heterogeneous public opinion network data related to each stage of the historical public opinion events that are already in the decline phase, based on one or more keywords associated with each historical event and according to a preset time threshold. For example, according to the development timeline of the historical public opinion events, starting from the incubation period, covering the formation period, fluctuation period, and decline period, the preset time threshold is set to 24 hours, and data is collected sequentially. The public opinion network data collected from multiple information sources on the Internet is divided into 24-hour units, and the public opinion network data from multiple information sources within each 24-hour period is regarded as a multi-source heterogeneous public opinion network data. Each multi-source heterogeneous public opinion network data is preprocessed to obtain preprocessed multi-source heterogeneous public opinion network data. The preprocessing of the acquired multi-source heterogeneous public opinion network data includes data cleaning and data integration. Data that is irrelevant to the public opinion events and invalid data are removed from the acquired multi-source heterogeneous public opinion network data, and then the cleaned data is integrated and summarized.

[0245] Continuing in this embodiment, the second module 220 of the device can perform feature quantization processing on each preprocessed multi-source heterogeneous public opinion network data based on a preset feature library, obtaining the feature value of each feature in the preset feature library corresponding to each preprocessed multi-source heterogeneous public opinion network data, which serves as the first feature set of each preprocessed multi-source heterogeneous public opinion network data. Specifically, based on professional knowledge and practical experience, combined with the analysis of different historical public opinion events, a feature library influencing the development of public opinion events can be pre-summarized. For each feature in this feature library, feature quantization processing is performed on each preprocessed multi-source heterogeneous public opinion network data, obtaining the feature value of each feature in the preset feature library corresponding to each preprocessed multi-source heterogeneous public opinion network data, which can serve as the first feature set of each preprocessed multi-source heterogeneous public opinion network data. The quantization method used is related to the data type of the public opinion network, and can be a statistical value, average value, or percentage quantization method, etc., which is not limited here.

[0246] Continuing in this embodiment, the third module 230 of the device performs feature selection on the first feature set of each preprocessed multi-source heterogeneous public opinion network data based on the Pearson correlation coefficient to obtain the second feature set of the preprocessed multi-source heterogeneous public opinion network data. Based on the second feature sets of all preprocessed multi-source heterogeneous public opinion network data, a third feature set of each preprocessed multi-source heterogeneous public opinion network data is obtained. Furthermore, the feature values ​​of each feature in the third feature set of each preprocessed public opinion network data are standardized to obtain a fourth feature set of each preprocessed public opinion network data. The third feature set of each preprocessed public opinion network data includes several identical features and their feature values.

[0247] Continuing in this embodiment, the fourth module 240 of the device can perform truth labeling on each preprocessed multi-source heterogeneous public opinion network data based on the stage of the historical public opinion event corresponding to each preprocessed multi-source heterogeneous public opinion network data, and take the fourth feature set and the corresponding truth value of each preprocessed multi-source heterogeneous public opinion network data as a data sample. The above operations are also performed on different historical public opinion events that are in the decline phase, and a sample dataset is formed based on all the obtained data samples.

[0248] Continuing in this embodiment, the fifth module 250 of the device can train a PNN neural network model based on the sample dataset. When the prediction accuracy of the trained PNN neural network model meets a preset threshold, the construction of the public opinion event fading prediction model is completed. Specifically, the PNN neural network model is trained based on the obtained sample dataset, and the construction of the public opinion event fading prediction model is completed when the prediction accuracy of the trained PNN neural network model meets a preset threshold.

[0249] Optionally, the device can also use the public opinion event receding prediction model to predict whether a public opinion event is in a receding phase, wherein, for example... Figure 2 An alternative embodiment of the apparatus shown further includes:

[0250] The sixth module 260 is used to acquire multi-source heterogeneous public opinion network data related to the public opinion event based on the preset time threshold, and to preprocess the multi-source heterogeneous public opinion network data to obtain preprocessed multi-source heterogeneous public opinion network data.

[0251] The seventh module 270 is used to perform feature quantization processing on the preprocessed multi-source heterogeneous public opinion network data based on the aforementioned features, to obtain feature values ​​corresponding to the aforementioned features, and to perform standardization processing on each feature value to obtain a feature set of the preprocessed multi-source heterogeneous public opinion network data, wherein the feature set includes the aforementioned features and their standardized feature values.

[0252] The eighth module 280 is used to input the feature set into the public opinion event decline prediction model, and to determine whether the public opinion event is in the decline period based on the output of the public opinion event decline prediction model.

[0253] In this optional embodiment, the sixth module 260 of the device can collect network information containing keywords or combinations of keywords related to the public opinion event based on one or more keywords associated with the public opinion event, using the same preset time threshold as when constructing the public opinion event decline prediction model, to obtain multi-source heterogeneous public opinion network data related to the public opinion event in units of the preset time threshold. For example, if the preset time threshold is 24 hours, multi-source heterogeneous public opinion network data related to the public opinion event can be obtained within 24 hours before the current time point or the time point to be monitored based on one or more keywords related to the public opinion event, and the obtained multi-source heterogeneous public opinion network data can be preprocessed to obtain preprocessed multi-source heterogeneous public opinion network data.

[0254] Continuing in this optional embodiment, the seventh module 270 of the device can perform feature quantization processing on the preprocessed multi-source heterogeneous public opinion network data based on the several features, obtain feature values ​​corresponding to the several features, and perform standardization processing on each feature value to obtain the feature set of the preprocessed multi-source heterogeneous public opinion network data, wherein the feature set includes the several features and their standardized feature values.

[0255] Continuing with this optional embodiment, the third module 280 of the device can input the several standardized feature values ​​into a pre-constructed public opinion event decline prediction model, and determine whether the public opinion event is in the decline phase based on the output of the public opinion event decline prediction model. Specifically, by inputting the feature set into the public opinion event decline prediction model, it predicts whether the public opinion event corresponding to the pre-processed multi-source heterogeneous public opinion network data is in the decline phase. Based on the output of the public opinion event decline model, it can be determined whether the public opinion event is in the decline phase.

[0256] In this optional embodiment, the above-mentioned modules of the device can collect public opinion network data related to public opinion events from multiple information sources within a preset time threshold at the current time point or the time point to be monitored. After processing, the feature set of the public opinion network data is obtained. The feature set is then input into the constructed public opinion event decline prediction model to predict whether the public opinion event is in the decline period.

[0257] In the various embodiments and / or optional embodiments of the above-described apparatus, the parts of the method steps performed by each module that are not mentioned in the above-described related method embodiments and / or optional embodiments are the same as those described in the above-described related method embodiments and / or optional embodiments, and will not be repeated here.

[0258] According to another aspect of this application, a computer-readable medium is also provided, the computer-readable medium storing computer-readable instructions that can be executed by a processor to implement the foregoing method embodiments.

[0259] It should be noted that the order in which the methods and / or optional embodiments in this application are performed is not strictly limited, as long as the methods and / or optional embodiments can solve the defects of the prior art, achieve the inventive purpose of this application, and obtain beneficial effects. The methods and / or optional embodiments in this application can be implemented in software and / or a combination of software and hardware. The software program involved in this application can be executed by a processor to implement the steps or functions of the above embodiments. Similarly, the software program of this application (including related data structures) can be stored in a computer-readable recording medium.

[0260] Furthermore, part or all of this application can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. The program instructions invoking the methods of this application may be stored in a fixed or removable recording medium, and / or transmitted via data streams in broadcast or other signal carrying media, and / or stored in the working memory of a computer device operating according to the program instructions.

[0261] According to another aspect of this application, an apparatus for constructing a predictive model for the fading of public opinion events is also provided. The apparatus includes: a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the apparatus is triggered to run the methods and / or technical solutions of the foregoing embodiments.

[0262] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims. Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in the apparatus claims may also be implemented by a single unit or device through software and / or hardware. The terms "first," "second," etc., are used to indicate names and do not indicate any particular order.

Claims

1. A method for predicting the fading of public opinion events, characterized in that, The method includes: Based on a preset time threshold, several multi-source heterogeneous public opinion network data at different stages related to historical public opinion events that are already in the receding phase are acquired, and each multi-source heterogeneous public opinion network data is preprocessed to obtain multiple preprocessed multi-source heterogeneous public opinion network data. Based on the preset feature library, feature quantization is performed on each preprocessed multi-source heterogeneous public opinion network data to obtain the feature value of each feature in the preset feature library corresponding to each preprocessed multi-source heterogeneous public opinion network data, which serves as the first feature set of each preprocessed multi-source heterogeneous public opinion network data. Based on the Pearson correlation coefficient, feature selection is performed on the first feature set of each preprocessed multi-source heterogeneous public opinion network data to obtain the second feature set of the preprocessed multi-source heterogeneous public opinion network data. Based on the second feature sets of all preprocessed multi-source heterogeneous public opinion network data, a third feature set of each preprocessed multi-source heterogeneous public opinion network data is obtained. The feature values ​​of each feature in the third feature set of each preprocessed public opinion network data are standardized to obtain a fourth feature set of each preprocessed public opinion network data. The third feature set of each preprocessed public opinion network data includes several identical features and their feature values. Based on the stage of the historical public opinion event corresponding to each preprocessed multi-source heterogeneous public opinion network data, the truth value is labeled for each preprocessed multi-source heterogeneous public opinion network data, and the fourth feature set and the corresponding truth value of each preprocessed multi-source heterogeneous public opinion network data are used as a data sample. The above operation is also performed on different historical public opinion events that are in the decline phase. Based on all the obtained data samples, a sample dataset is formed. Based on the sample dataset, a PNN neural network model is trained. When the prediction accuracy of the trained PNN neural network model meets a preset threshold, the construction of the public opinion event fading prediction model is completed.

2. The method according to claim 1, characterized in that, The multi-source heterogeneous public opinion network data includes: Data related to the aforementioned public opinion event from different information sources on the Internet, wherein the data includes textual and non-textual data.

3. The method according to claim 1, characterized in that, The aforementioned features include: Features under several multi-dimensional dimensions, wherein the multi-dimensional dimensions include multiple first-level dimensions, each first-level dimension includes multiple second-level dimensions, and each second-level dimension includes several features.

4. The method according to claim 3, characterized in that, The multiple primary dimensions include: Disseminator; Content to be disseminated; Motivation for dissemination; Propagation and evolution.

5. The method according to claim 4, characterized in that, The secondary dimensions under the communicator include: Media participation; Influence of self-media; Credibility of identity.

6. The method according to claim 5, characterized in that, Several characteristics of media engagement include: The amount of media coverage at the central level; Provincial media coverage; The amount of media coverage at the prefecture-level city level; Other media coverage.

7. The method according to claim 5, characterized in that, Several characteristics of the influence of self-media include: Percentage of original content created by personally verified users; Percentage of reposts by personally verified users; Percentage of comments from personally verified users; Percentage of shares received by personally verified users; Percentage of comments received by personally verified users; Percentage of likes received by personally verified users.

8. The method according to claim 5, characterized in that, The features under the identity credibility include: The proportion of information about the location where the public opinion event occurred.

9. The method according to claim 4, characterized in that, The secondary dimensions under the content being disseminated include: Fineness; Sentimental inclination; Richness.

10. The method according to claim 9, characterized in that, The features at this level of detail include: Average word count per original post; Average number of characters per forwarded message; Average word count per comment.

11. The method according to claim 9, characterized in that, Several characteristics under the sentiment tendency include: Positive original content and reposts; Neutral original content and repost volume; Negative original content and reposts; Number of positive comments; Neutral comment volume; Number of negative comments.

12. The method according to claim 9, characterized in that, Several features under the richness level include: Only the proportion of information in short links is included; This only includes the proportion of information in the video. Only the information content of the images is included; This includes the proportion of information in short links and videos; The proportion of information including short links and images; This includes the proportion of information presented in videos and images; This includes the proportion of information in short links, videos, and images; This includes the proportion of information in videos.

13. The method according to claim 4, characterized in that, The secondary dimensions under the aforementioned motivation for dissemination include: Motivation intensity; Motivational expression level.

14. The method according to claim 13, characterized in that, Several characteristics under the aforementioned motivational intensity include: The proportion of extreme information.

15. The method according to claim 13, characterized in that, Several features under the motivational expression level include: Degree of behavioral intention expression; Deep causal expression level; The continued leaks and high level of attention; Public attention to the outcome of the investigation.

16. The method according to claim 4, characterized in that, The secondary dimensions under propagation and evolution include: Dissemination; Evolutionary degree.

17. The method according to claim 16, characterized in that, Several characteristics under the propagation degree include: Original information content; The amount of information forwarded; Information content in the comments.

18. The method according to claim 16, characterized in that, Several features under the aforementioned degree of evolution include: The rate of change in information volume compared to the previous period.

19. The method according to claim 1, characterized in that, The method further includes: Based on the preset time threshold, multi-source heterogeneous public opinion network data related to the public opinion event is acquired, and the multi-source heterogeneous public opinion network data is preprocessed to obtain preprocessed multi-source heterogeneous public opinion network data. Based on the aforementioned features, feature quantization processing is performed on the preprocessed multi-source heterogeneous public opinion network data to obtain feature values ​​corresponding to the aforementioned features. Then, the feature values ​​of each feature are standardized to obtain the feature set of the preprocessed multi-source heterogeneous public opinion network data. The feature set includes the aforementioned features and their standardized feature values. The feature set is input into the public opinion event decline prediction model, and based on the output of the public opinion event decline prediction model, it is determined whether the public opinion event is in the decline period.

20. An apparatus for constructing a predictive model for the fading of public opinion events, characterized in that, The device includes: The first module is used to acquire several multi-source heterogeneous public opinion network data at different stages related to historical public opinion events that are already in the receding period, based on a preset time threshold, and to preprocess each multi-source heterogeneous public opinion network data to obtain multiple preprocessed multi-source heterogeneous public opinion network data. The second module is used to perform feature quantization processing on each preprocessed multi-source heterogeneous public opinion network data based on a preset feature library, to obtain the feature value of each feature in the preset feature library corresponding to each preprocessed multi-source heterogeneous public opinion network data, which serves as the first feature set of each preprocessed multi-source heterogeneous public opinion network data. The third module is used to perform feature selection on the first feature set of each preprocessed multi-source heterogeneous public opinion network data based on the Pearson correlation coefficient to obtain the second feature set of the preprocessed multi-source heterogeneous public opinion network data, and to obtain the third feature set of each preprocessed multi-source heterogeneous public opinion network data based on the second feature set of all preprocessed multi-source heterogeneous public opinion network data, and to standardize the feature values ​​of each feature in the third feature set of each preprocessed public opinion network data to obtain the fourth feature set of each preprocessed public opinion network data, wherein the third feature set of each preprocessed public opinion network data includes the same number of features and their feature values; The fourth module is used to perform truth labeling on each preprocessed multi-source heterogeneous public opinion network data based on the stage of the historical public opinion event corresponding to each preprocessed multi-source heterogeneous public opinion network data, and to take the fourth feature set and the corresponding truth value of each preprocessed multi-source heterogeneous public opinion network data as a data sample. The above operation is also performed on different historical public opinion events that are in the decline phase. Based on all the obtained data samples, a sample dataset is formed. The fifth module is used to train a PNN neural network model based on the sample dataset. When the prediction accuracy of the trained PNN neural network model meets the preset threshold, the construction of the public opinion event fading prediction model is completed.

21. The apparatus according to claim 20, characterized in that, The device further includes: The sixth module is used to acquire multi-source heterogeneous public opinion network data related to the public opinion event based on the preset time threshold, and to preprocess the multi-source heterogeneous public opinion network data to obtain preprocessed multi-source heterogeneous public opinion network data. The seventh module is used to perform feature quantization processing on the preprocessed multi-source heterogeneous public opinion network data based on the aforementioned features, to obtain feature values ​​corresponding to the aforementioned features, and to perform standardization processing on each feature value to obtain the feature set of the preprocessed multi-source heterogeneous public opinion network data, wherein the feature set includes the aforementioned features and their standardized feature values. The eighth module is used to input the feature set into the public opinion event decline prediction model, and to determine whether the public opinion event is in the decline period based on the output of the public opinion event decline prediction model.

22. A computer-readable medium, characterized in that, It stores computer-readable instructions that are executed by a processor to implement the method as described in any one of claims 1 to 19.

23. A device for constructing a predictive model for the fading of public opinion events, characterized in that, The device includes: One or more processors; and A memory storing computer-readable instructions, which, when executed, cause the processor to perform the operations of the method as described in any one of claims 1 to 19.

Images