CN117576504A - Training method of social media false news detection model - Google Patents

Training method of social media false news detection model Download PDF

Info

Publication number
CN117576504A
CN117576504A CN202311475479.XA CN202311475479A CN117576504A CN 117576504 A CN117576504 A CN 117576504A CN 202311475479 A CN202311475479 A CN 202311475479A CN 117576504 A CN117576504 A CN 117576504A
Authority
CN
China
Prior art keywords
news
social media
propagation
field
false
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202311475479.XA
Other languages
Chinese (zh)
Inventor
张怀文
杨青
刘鑫鑫
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Inner Mongolia University
Original Assignee
Inner Mongolia University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Inner Mongolia University filed Critical Inner Mongolia University
Priority to CN202311475479.XA priority Critical patent/CN117576504A/en
Publication of CN117576504A publication Critical patent/CN117576504A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/042—Knowledge-based neural networks; Logical representations of neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/0895—Weakly supervised learning, e.g. semi-supervised or self-supervised learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/778—Active pattern-learning, e.g. online learning of image or video features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/95—Pattern authentication; Markers therefor; Forgery detection

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Multimedia (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention provides a training method of a social media false news detection model, which comprises the following steps: acquiring a plurality of news propagation graphs of a first field in a training set; the news propagation graph is used for representing a news propagation path; acquiring a plurality of news propagation graphs of a second field in the training set; the ratio of the number of news maps of the second domain to the number of news maps of the first domain is less than a threshold; training the social media false news detection model according to the plurality of news propagation graphs in the first field, the plurality of news propagation graphs in the second field and the first target loss function to obtain a social media false news detection model trained based on the training set; the social media false news detection model is used for detecting the authenticity of news; the first objective loss function is determined by the classification loss, the global contrast loss, and the local contrast loss. The method of the invention realizes the accuracy of false news detection in the second field with less data quantity and shorter propagation time.

Description

Training method of social media false news detection model
Technical Field
The invention relates to the technical field of data processing, in particular to training of a social media false news detection model.
Background
With the rapid development of the internet, social media becomes an important platform for people to acquire information, publish ideas and communicate daily. However, with the increase in the number of users of social platforms, if a false news is a trending topic, social stability will be affected and potential economic loss risks will be brought through the discussion and propagation of a large number of users. In order to determine the authenticity of news, people propose a social media false news detection task aimed at determining the authenticity of news in social media.
In the related art, an automatic false news detection model trained in a high-resource domain with a large amount of data can accurately detect the high-resource domain; however, for the emerging field of emergency generation, the accuracy of the automatic false news detection effect is lower due to insufficient data. Therefore, how to solve the problem that the news in the low resource domain is actually detected accurately under the condition that the high resource data is sufficient and the low resource data is less is a urgent need for those skilled in the art.
Disclosure of Invention
Aiming at the problems in the prior art, the embodiment of the invention provides a training method of a social media false news detection model.
Specifically, the embodiment of the invention provides the following technical scheme:
In a first aspect, an embodiment of the present invention provides a training method for a social media false news detection model, including:
acquiring a plurality of news propagation graphs of a first field in a training set; the news propagation graph is used for representing a news propagation path;
acquiring a plurality of news propagation graphs of a second field in the training set; the ratio of the number of news maps of the second domain to the number of news maps of the first domain is less than a threshold;
training the social media false news detection model according to the plurality of news propagation graphs in the first field, the plurality of news propagation graphs in the second field and the first target loss function to obtain a social media false news detection model trained based on the training set; the social media false news detection model is used for detecting the authenticity of news; the first objective loss function is determined from the classification loss, the global contrast loss, and the local contrast loss; the global contrast loss represents the node characteristics in the news propagation graph, and the degree of association between the node characteristics in the first type of the news propagation graph and the news propagation graph characteristics; the local contrast loss represents a degree of association between node features in a second type of augmentation graph of the news propagation graph and node features in a third type of augmentation graph of the news propagation graph; the classification loss indicates the accuracy of the classification result.
Further, the social media false news detection model comprises at least one of the following:
a feature extraction module; the feature extraction module is used for extracting news propagation graph features;
a classification module; the classification module is used for predicting the authenticity of social media news corresponding to the news propagation graph according to the news propagation graph characteristics;
a self-supervision learning module; the self-supervision learning module is used for determining global contrast loss according to the node characteristics in the news propagation graph, the node characteristics in the first type of the news propagation graph and the news propagation graph characteristics; and determining local contrast loss according to the node characteristics in the second type augmentation chart of the news propagation chart and the node characteristics in the third type augmentation chart of the news propagation chart.
Further, the social media false news detection model is trained based on the following:
inputting a plurality of news propagation graphs of a first field and a plurality of news propagation graphs of a second field in a training set into a social media false news detection model, and outputting an authenticity detection result of social media news corresponding to the news propagation graphs; obtaining classification loss of the social media false news detection model according to the authenticity detection result of the social media news and the label information of the news; the tag information is used for marking the authenticity of news;
Inputting a plurality of news propagation graphs in the first field, a first type of augmentation graph corresponding to the plurality of news propagation graphs in the first field, a second type of augmentation graph corresponding to the plurality of news propagation graphs in the first field and a third type of augmentation graph corresponding to the plurality of news propagation graphs in the first field into a social media false news detection model to obtain global contrast loss and local contrast loss of a plurality of news in the first field;
inputting a plurality of news propagation graphs in the second field in the training set, a first type augmentation graph corresponding to the plurality of news propagation graphs in the second field, a second type augmentation graph corresponding to the plurality of news propagation graphs in the second field and a third type augmentation graph corresponding to the plurality of news propagation graphs in the second field into a social media false news detection model to obtain global contrast loss and local contrast loss of a plurality of news in the second field;
taking the weighted sum of the classification loss of the social media false news detection model, the global contrast loss and the local contrast loss of the news in the first field, and the global contrast loss and the local contrast loss of the news in the second field as the value of a first target loss function;
Training the social media false news detection model based on the value of the first target loss function to obtain the social media false news detection model trained based on the training set.
Further, the social media false news detection model further comprises:
a data self-adaptive constraint module; the data adaptive constraint module is used for determining differences between the news spread map features in the training set and the news spread map features in the test set.
Further, training the social media false news detection model based on the value of the first target loss function, and after obtaining the social media false news detection model trained based on the training set, further comprising:
inputting a plurality of news propagation graphs in the second field in the test set, a first type of augmentation graph corresponding to the plurality of news propagation graphs in the second field in the test set, a second type of augmentation graph corresponding to the plurality of news propagation graphs in the second field in the test set and a third type of augmentation graph corresponding to the plurality of news propagation graphs in the second field in the test set to a social media false news detection model trained based on the training set to obtain global contrast loss and local contrast loss of a plurality of news in the second field in the test set;
Taking the sum of the global contrast loss and the local contrast loss weights of a plurality of news in the second field in the test set as a value of a second target loss function according to the difference between the news propagation map features in the second field in the training set and the news propagation map features in the second field in the test set;
training the social media false news detection model based on the value of the second target loss function to obtain the social media false news detection model based on the potential characteristics of the test set.
In a second aspect, an embodiment of the present invention further provides a social media false news detection method, including:
acquiring a news propagation diagram of a second field to be detected;
inputting the news propagation diagram of the second field to be detected into a social media false news detection model to obtain an authenticity detection result of social media news corresponding to the news propagation diagram of the second field; the social media false news detection model is trained based on the training method of the social media false news detection model as the first aspect.
In a third aspect, an embodiment of the present invention further provides a social media false news detection device, including:
acquiring a news propagation diagram of a second field to be detected;
Inputting the news propagation diagram of the second field to be detected into a social media false news detection model to obtain an authenticity detection result of social media news corresponding to the news propagation diagram of the second field; the social media false news detection model is trained based on the training method of the social media false news detection model according to the first aspect.
In a fourth aspect, an embodiment of the present invention further provides an electronic device, including a memory, a processor, and a computer program stored on the memory and capable of running on the processor, where the processor implements the training method of the social media false news detection model according to the first aspect or the social media false news detection method according to the second aspect when executing the program.
In a fifth aspect, an embodiment of the present invention further provides a non-transitory computer readable storage medium, on which a computer program is stored, where the computer program is executed by a processor to implement the training method of the social media false news detection model according to the first aspect or the social media false news detection method according to the second aspect.
In a sixth aspect, an embodiment of the present invention further provides a computer program product, including a computer program, where the computer program when executed by a processor implements the training method of the social media false news detection model according to the first aspect or the social media false news detection method according to the second aspect.
According to the training method for the social media false news detection model, the training is carried out on the social media false news detection model according to the plurality of news propagation diagrams in the first field, the plurality of news propagation diagrams in the second field and the first target loss function, so that the trained social media false news detection model not only can accurately extract characteristic information of the news in the first field and global contrast loss between nodes in the news propagation diagrams and local contrast loss between nodes in the augmented diagrams in the news propagation diagrams, but also can accurately extract characteristic information of the news in the second field and global contrast loss between nodes in the news propagation diagrams and local contrast loss between nodes in the augmented diagrams in the news propagation diagrams, and therefore the trained social media false news detection model can accurately and comprehensively extract the characteristic information of the false news in the second field with less data quantity and shorter propagation time, the detection accuracy of the false news in the second field is effectively promoted and improved, and the trained social media false news detection model can accurately extract the real news in the second field with less data quantity and shorter propagation time.
Drawings
In order to more clearly illustrate the invention or the technical solutions of the prior art, the following description will briefly explain the drawings used in the embodiments or the description of the prior art, and it is obvious that the drawings in the following description are some embodiments of the invention, and other drawings can be obtained according to the drawings without inventive effort for a person skilled in the art.
FIG. 1 is a schematic flow diagram of a training method of a social media false news detection model according to an embodiment of the present invention;
FIG. 2 is a second flowchart of a training method of a social media false news detection model according to an embodiment of the present invention;
FIG. 3 is a third flow chart of a training method of a social media false news detection model according to an embodiment of the present invention;
FIG. 4 is a flowchart of a training method of a social media false news detection model according to an embodiment of the present invention;
FIG. 5 is a schematic structural diagram of a training device of a social media false news detection model according to an embodiment of the present invention;
fig. 6 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is apparent that the described embodiments are some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The method provided by the embodiment of the invention can be applied to the false news detection scene, and improves the accuracy of the false news detection in the second field with less data quantity and shorter propagation time.
In the related art, an automatic false news detection model trained in a high-resource domain with a large amount of data can accurately detect the high-resource domain; however, for the emerging field of emergency generation, the accuracy of the automatic false news detection effect is lower due to insufficient data. Therefore, how to solve the problem that the news in the low resource domain is actually detected accurately under the condition that the high resource data is sufficient and the low resource data is less is a urgent need for those skilled in the art.
According to the training method of the social media false news detection model, the social media false news detection model is trained according to the plurality of news propagation diagrams in the first field, the plurality of news propagation diagrams in the second field and the first target loss function, so that the trained social media false news detection model not only can accurately extract characteristic information of the news in the first field and global contrast loss between nodes in the news propagation diagrams and local contrast loss between nodes in the augmented diagrams in the news propagation diagrams, but also can accurately extract characteristic information of the news in the second field and global contrast loss between nodes in the news propagation diagrams and local contrast loss between nodes in the augmented diagrams in the news propagation diagrams, and therefore the trained social media false news detection model can accurately and comprehensively extract the characteristic information of the false news in the second field with less data volume and shorter propagation time, the detection accuracy of the false news in the second field is effectively promoted and improved, and the trained social media false news detection model can accurately detect the false news in the second field with less data volume and the real time.
The following describes the technical scheme of the present invention in detail with reference to fig. 1 to 6. The following embodiments may be combined with each other, and some embodiments may not be repeated for the same or similar concepts or processes.
Fig. 1 is a flowchart of an embodiment of a training method of a social media false news detection model according to an embodiment of the present invention. As shown in fig. 1, the method provided in this embodiment includes:
step 101, acquiring a plurality of news propagation graphs of a first field in a training set; the news propagation graph is used for representing a news propagation path;
specifically, the efficiency of judging and detecting the authenticity of news based on subjective consciousness of people in the prior art is low. As shown in fig. 2, the domain with relatively large data size is the source domain, the domain with relatively small data size is the target domain, the false news of the first domain can be accurately identified (only one error) through the model trained by the source domain and the target domain, and a large number of errors (6 errors) exist in the detection result of the false news of the second domain.
In order to solve the above problem, in the embodiment of the present application, a plurality of news propagation graphs in a first field in a training set are first acquired; the news propagation graph is used for representing a news propagation path; optionally, the news in the social media is propagated mainly by the behavior of the user in the social media, and the user downloads, comments or praise the news in the social media, so that a news propagation diagram of the social media is formed; that is, the user forwards, comments or praise the news as the nodes in the news propagation diagram, and the news propagation diagram can be formed according to the nodes and time information in the news propagation diagram, so that the news propagation path is accurately represented.
102, acquiring a plurality of news propagation graphs in a second field in a training set; the ratio of the number of news maps of the second domain to the number of news maps of the first domain is less than a threshold;
specifically, when an emergency occurs, that is, a new domain occurs, how to accurately detect the authenticity of the news in the new domain under the condition of less data, so that measures are timely taken to avoid further propagation of false news, which is a problem to be solved by those skilled in the art. In the embodiment of the application, after the plurality of news dissemination maps of the first field are acquired, a plurality of news dissemination maps of a second field in the training set are also acquired, wherein the number of the news dissemination maps of the second field is far smaller than that of the news dissemination maps of the first field; for example, the ratio of the number of news maps of the second domain to the number of news maps of the first domain is less than a threshold; for example, the number of news maps of the second domain is 10% of the number of news maps of the first domain; in the training process of the social media false news detection model, news in the first field with more data quantity and longer propagation time is obtained as a training sample, and news in the second field with less data quantity and shorter propagation time is also obtained as a training sample, so that the trained social media false news detection model can accurately detect the authenticity of the news with more data quantity and longer propagation time, and can accurately detect the authenticity of the social media news with less data quantity and shorter propagation time.
Step 103, training the social media false news detection model according to the plurality of news propagation graphs in the first field, the plurality of news propagation graphs in the second field and the first target loss function to obtain a social media false news detection model trained based on the training set; the social media false news detection model is used for detecting the authenticity of news; the first objective loss function is determined from the classification loss, the global contrast loss, and the local contrast loss; the global contrast loss represents the node characteristics in the news propagation graph, and the degree of association between the node characteristics in the first type of the news propagation graph and the news propagation graph characteristics; the local contrast loss represents a degree of association between node features in a second type of augmentation graph of the news propagation graph and node features in a third type of augmentation graph of the news propagation graph; the classification loss indicates the accuracy of the classification result.
Specifically, in the embodiment of the application, after a plurality of news propagation graphs of a first field with a large data volume and a long propagation time and a news propagation graph of a second field with a small data volume and a short propagation time are obtained, training a social media false news detection model according to the plurality of news propagation graphs of the first field, the plurality of news propagation graphs of the second field and a first target loss function to obtain a social media false news detection model trained based on a training set; the social media false news detection model is used for detecting the authenticity of news. Optionally, the first objective loss function is determined from a classification loss, a global contrast loss, and a local contrast loss; the global contrast loss is used for representing node characteristics in the news propagation graph, and the degree of association between the node characteristics in the first type of the news propagation graph and the news propagation graph characteristics; the local contrast loss is used to represent a degree of association between node features in a second type of augmentation graph of the news feed graph and node features in a third type of augmentation graph of the news feed graph; that is, global contrast loss is used to represent the degree of association of a node with a news map, and local contrast loss is used to represent the degree of association between nodes. Optionally, the global contrast loss can be used for carrying out contrast learning according to the original news propagation chart and an augmented chart obtained by randomly disturbing the nodes of the original news propagation chart, so as to obtain the relationship characteristics between each node in the original news propagation chart and the whole chart; the local contrast loss can be based on two kinds of augmentation graphs obtained by selectively discarding edges of the original image of the news propagation graph and selectively covering node features of the original image, and the relationship features between each node in the graph obtained by contrast learning. In other words, in the embodiment of the application, the global contrast loss and the local contrast loss are increased, so that the feature extractor in the social media false news detection model is optimized, the features in the news propagation graph can be extracted more accurately and comprehensively by the optimized feature extractor and the social media false news detection model, the feature extraction of the false news is more accurate and complete, more important information for news classification can be extracted, and better and more accurate classification of the authenticity of the news can be realized. On the other hand, according to the training method of the social media false news detection model, local contrast learning and global contrast learning are carried out on the second-field false news data and the augmented data thereof, so that the local information and the global information of the second-field false news data are learned, and the generalization capability of the model on the second-field false news data is improved.
After the social media false news detection model is trained based on the first target loss function, the trained social media false news detection model not only can accurately extract the characteristic information of the first field news, the global contrast loss between the nodes in the news propagation graph and the local contrast loss between the augmented graph nodes in the news propagation graph, but also can accurately extract the characteristic information of the second field news, the global contrast loss between the nodes in the news propagation graph and the augmented graph nodes in the news propagation graph and the local contrast loss between the augmented graph nodes in the news propagation graph, so that the social media false news detection model can accurately identify and extract the characteristic information of the false news, and effectively promote and improve the detection accuracy of the false news, thereby also enabling the trained social media false news detection model to accurately detect the news in the second field with less data quantity and shorter propagation time, and improving the authenticity of the false news in the second field with less data quantity and shorter propagation time.
According to the method, according to the plurality of news propagation graphs in the first field, the plurality of news propagation graphs in the second field and the first objective loss function, the trained social media false news detection model is trained, so that the trained social media false news detection model not only can accurately extract the characteristic information of the news in the first field and the global contrast loss between the nodes in the news propagation graphs and the local contrast loss between the nodes in the news propagation graphs, but also can accurately extract the characteristic information of the news in the second field and the global contrast loss between the nodes in the news propagation graphs and the local contrast loss between the nodes in the news propagation graphs, so that the trained social media false news detection model can accurately and comprehensively extract the characteristic information of the false news in the second field with less data quantity and shorter propagation time, the detection accuracy of the false news is effectively promoted and improved, and the trained social media false news detection model can accurately detect the false news in the second field with less data quantity and shorter propagation time, and the true false news detection accuracy of the second field with less data quantity and shorter propagation time is improved.
In one embodiment, the social media false news detection model includes at least one of:
a feature extraction module; the feature extraction module is used for extracting news propagation graph features;
a classification module; the classification module is used for predicting the authenticity of social media news corresponding to the news propagation graph according to the news propagation graph characteristics;
a self-supervision learning module; the self-supervision learning module is used for determining global contrast loss according to the node characteristics in the news propagation graph, the node characteristics in the first type of the news propagation graph and the news propagation graph characteristics; and determining local contrast loss according to the node characteristics in the second type augmentation chart of the news propagation chart and the node characteristics in the third type augmentation chart of the news propagation chart.
Specifically, in the embodiment of the application, the social media false news detection model comprises a feature extraction module, a classification module and a self-supervision learning module; the feature extraction module is used for extracting news propagation graph features; optionally, the feature extraction module is built based on a graph rolling network GCN; the classification module is used for predicting the authenticity of social media news corresponding to the news propagation graph according to the news propagation graph characteristics; the self-supervision learning module is used for determining global contrast loss according to the node characteristics in the news propagation graph, the node characteristics in the first type of the news propagation graph and the news propagation graph characteristics; according to the node characteristics in the second type of the news propagation graph and the node characteristics in the third type of the news propagation graph, determining local contrast loss, so that the social media false news detection model not only can accurately extract the characteristic information of the first field news, the global contrast loss between the nodes in the news propagation graph and the news propagation graph, the local contrast loss between the nodes in the news propagation graph, but also can accurately extract the characteristic information of the second field news, the global contrast loss between the nodes in the news propagation graph and the local contrast loss between the nodes in the news propagation graph; optionally, the global contrast loss and the local contrast loss corresponding to the false news and the global contrast loss and the local contrast loss corresponding to the true news are different; the global contrast loss and the local contrast loss corresponding to the first domain false news and the global contrast loss and the local contrast loss corresponding to the second domain false news are also different, so that the trained social media false news detection model can accurately identify and extract the characteristic information of the false news, the detection of the authenticity of the second domain news with smaller data volume and shorter propagation time is realized, and the accuracy of the detection of the second domain false news with smaller data volume and shorter propagation time is improved.
In one embodiment, the social media false news detection model is trained based on the following:
inputting a plurality of news propagation graphs of a first field and a plurality of news propagation graphs of a second field in a training set into a social media false news detection model, and outputting an authenticity detection result of social media news corresponding to the news propagation graphs; obtaining classification loss of the social media false news detection model according to the authenticity detection result of the social media news and the label information of the news; the tag information is used for marking the authenticity of news;
inputting a plurality of news propagation graphs in the first field, a first type of augmentation graph corresponding to the plurality of news propagation graphs in the first field, a second type of augmentation graph corresponding to the plurality of news propagation graphs in the first field and a third type of augmentation graph corresponding to the plurality of news propagation graphs in the first field into a social media false news detection model to obtain global contrast loss and local contrast loss of a plurality of news in the first field;
inputting a plurality of news propagation graphs in the second field in the training set, a first type augmentation graph corresponding to the plurality of news propagation graphs in the second field, a second type augmentation graph corresponding to the plurality of news propagation graphs in the second field and a third type augmentation graph corresponding to the plurality of news propagation graphs in the second field into a social media false news detection model to obtain global contrast loss and local contrast loss of a plurality of news in the second field;
Taking the weighted sum of the classification loss of the social media false news detection model, the global contrast loss and the local contrast loss of the news in the first field, and the global contrast loss and the local contrast loss of the news in the second field as the value of a first target loss function;
training the social media false news detection model based on the value of the first target loss function to obtain the social media false news detection model trained based on the training set.
Specifically, in the training process of the social media false news detection model, a plurality of news propagation diagrams in a first field and a plurality of news propagation diagrams in a second field in a training set are input into the social media false news detection model, a classification module of the social media false news detection model outputs an authenticity detection result of social media news corresponding to the news propagation diagram, and then classification loss (detection accuracy) of the social media false news detection model can be obtained according to the authenticity detection result of the social media news and label information of the news.
Further, in the embodiment of the present application, a plurality of news propagation graphs in a first domain, a first type of augmentation graph corresponding to the plurality of news propagation graphs in the first domain, a second type of augmentation graph corresponding to the plurality of news propagation graphs in the first domain, and a third type of augmentation graph corresponding to the plurality of news propagation graphs in the first domain in a training set are input to a social media false news detection model, so as to obtain global contrast loss and local contrast loss of a plurality of news in the first domain; the self-supervision learning module in the social media false news detection model obtains local contrast loss and global contrast loss of nodes in the graph according to the characteristics of the original news propagation graph and three types of augmented graphs of the original news propagation graph; the global contrast loss is obtained by contrast learning according to a news propagation chart original image and an augmented chart obtained by randomly disturbing nodes of the news propagation chart original image, so that relationship characteristics between each node in the news propagation chart original image and the whole chart are obtained; the local contrast loss is based on the relationship characteristics between each node in the graph obtained by selectively discarding the edges of the original graph of the news propagation graph and selectively covering the node characteristics of the original graph.
For example, the global contrast loss between each node in the newscast artwork and the entire graph is modeled using the following formula:
D(Z si ,s)＝Sigmoid(Z si *s)
wherein Z is si Node feature representations representing news propagation graphs; s represents a feature representation of a news propagation map; * Representing the inner product; d represents a discriminator, and calculates the correlation scores of the positive and negative samples respectively; sigmoid is an activation function.
The global contrast loss between each node in the augmented graph obtained by randomly scrambling the nodes of the original graph of the news propagation graph and the whole graph is determined based on the following formula:
wherein N represents the number of nodes in the input graph; z is Z 0i Representing nodes in the original news propagation chart; z is Z 1i And the nodes in the augmented graph obtained by randomly scrambling the nodes of the original news propagation graph are represented.
Optionally, the local contrast loss is modeled using the following formula:
wherein Z is 2 ,Z 3 Two types of augmentation graphs representing news data; alternatively, the method can be two kinds of augmentation charts obtained by selectively discarding edges of original news propagation chart and selectively covering node features of the original chart;
(Z 2i ,Z 3j ) (i, j ε {1,...N }, i+.j), N represents the number of nodes in the graph, where cos () represents a cosine similarity function, τ is a superparameter and g () is used to further enhance the representation capabilities of the model. Representing the nodes calculated by g () ' as Z2' and Z3';
Where I represents the identity matrix, the local contrast loss of the node is ultimately defined as follows:
further, inputting a plurality of news propagation graphs in the second field, a first type of augmentation graph corresponding to the plurality of news propagation graphs in the second field, a second type of augmentation graph corresponding to the plurality of news propagation graphs in the second field and a third type of augmentation graph corresponding to the plurality of news propagation graphs in the second field into a social media false news detection model to obtain global contrast loss and local contrast loss of a plurality of news in the second field; optionally, the global contrast loss and the local contrast loss of the plurality of news in the second area are consistent with the determining method of the global contrast loss and the local contrast loss in the first area, which is not described in detail in the embodiment of the present application.
Optionally, after determining the classification loss of the social media false news detection model, the global contrast loss and the local contrast loss of the plurality of news in the first domain, and the global contrast loss and the local contrast loss of the plurality of news in the second domain, in the embodiment of the present application, the weighted sum of the classification loss of the social media false news detection model, the global contrast loss and the local contrast loss of the plurality of news in the first domain, and the global contrast loss and the local contrast loss of the plurality of news in the second domain is used as the value of the first target loss function; in other words, in the training process of the social media false news detection model, not only the classification loss of the social media false news detection model is simply considered, but also whether the social media false news detection model can accurately and comprehensively extract the characteristic information of the false news in the second field with smaller data volume and shorter propagation time is considered, so that the identification accuracy of the false news can be effectively promoted after the characteristic information of the false news in the second field with smaller data volume and shorter propagation time is accurately extracted, the problem that the characteristic information of the false news in the second field with smaller data volume and shorter propagation time cannot be accurately and comprehensively extracted is fundamentally solved, and the accuracy of the false news detection in the second field with smaller data volume and shorter propagation time is effectively improved.
According to the method, in the process of training the social media false news detection model, not only is the classification loss of the social media false news detection model simply considered, but also whether the social media false news detection model can accurately and comprehensively extract the characteristic information of the false news in the second field with smaller data quantity and shorter propagation time is considered, so that after the characteristic information of the false news in the second field with smaller data quantity and shorter propagation time is accurately extracted, the identification accuracy of the false news can be effectively promoted, the problem that the characteristic information of the false news in the second field with smaller data quantity and shorter propagation time cannot be accurately and comprehensively extracted is fundamentally solved, and the accuracy of the false news detection in the second field with smaller data quantity and shorter propagation time is effectively improved.
In an embodiment, the social media false news detection model further comprises:
a data self-adaptive constraint module; the data adaptive constraint module is used for determining differences between the news spread map features in the training set and the news spread map features in the test set.
Specifically, in the embodiment of the application, the social media false news detection model further comprises a data self-adaptive constraint module on the basis of comprising a feature extraction module, a classification module and a self-supervision learning module. The data self-adaptive constraint module is used for determining differences between the news propagation graph features in the training set and the news propagation graph features in the test set. That is, after training of the social media false news detection model is completed based on the training set, the difference between the news propagation graph feature in the training set and the news propagation graph feature in the test set is determined based on the data self-adaptive constraint module in the embodiment of the application, so that fine adjustment of the social media false news detection model is achieved, the feature extractor on the test set and the feature extractor on the training set are similar in performance, optimization of the feature extractor is achieved, the occurrence of the fitting problem is prevented, and therefore the social media false news detection model after the difference between the news propagation graph feature in the training set and the news propagation graph feature in the test set is constrained by the data self-adaptive constraint module can more accurately identify false news in the second field to be detected.
For example, the difference between the news map features of the second domain in the training set and the news map features of the second domain in the test set may be determined based on:
wherein h is the feature matrix of the training set data, mu represents the feature average value, sigma represents the covariance matrix, and the feature average value mu of the test set is obtained by the same method t Sum covariance matrix Σ t The differences between the news map features of the second domain in the training set and the news map features of the second domain in the test set are as follows:
the data self-adaptive constraint module in the embodiment of the application ensures that the social media false news detection model cannot be over-fitted to the false news data by calculating the difference between two statistical values of the false news data in the training set and the false news data in the test set, and improves the generalization capability of the model.
According to the method, the social media false news detection model is used for restraining the difference between the news propagation graph characteristics in the training set and the news propagation graph characteristics in the test set by adding the data self-adaptive restraining module, so that fine adjustment of the social media false news detection model is achieved, the feature extractor on the test set and the feature extractor on the training set are similar in performance, optimization of the feature extractor is achieved, the occurrence of the overfitting problem is prevented, and therefore the social media false news detection model after the difference between the news propagation graph characteristics in the training set and the news propagation graph characteristics in the test set is restrained by adding the data self-adaptive restraining module can be used for more accurately identifying false news in the second field to be detected.
In an embodiment, training the social media false news detection model based on the value of the first objective loss function, and after obtaining the social media false news detection model trained based on the training set, further includes:
inputting a plurality of news propagation graphs in the second field in the test set, a first type of augmentation graph corresponding to the plurality of news propagation graphs in the second field in the test set, a second type of augmentation graph corresponding to the plurality of news propagation graphs in the second field in the test set and a third type of augmentation graph corresponding to the plurality of news propagation graphs in the second field in the test set to a social media false news detection model trained based on the training set to obtain global contrast loss and local contrast loss of a plurality of news in the second field in the test set;
taking the sum of the global contrast loss and the local contrast loss weights of a plurality of news in the second field in the test set as a value of a second target loss function according to the difference between the news propagation map features in the second field in the training set and the news propagation map features in the second field in the test set;
training the social media false news detection model based on the value of the second target loss function to obtain the social media false news detection model based on the potential characteristics of the test set.
Specifically, in the embodiment of the present application, a plurality of news propagation graphs in a second field in a test set, a first type of augmentation graph corresponding to the plurality of news propagation graphs in the second field in the test set, a second type of augmentation graph corresponding to the plurality of news propagation graphs in the second field in the test set, and a third type of augmentation graph corresponding to the plurality of news propagation graphs in the second field in the test set are input into a social media false news detection model trained based on a training set, so as to obtain global contrast loss and local contrast loss of a plurality of news in the second field in the test set; optionally, the method steps of determining the global contrast loss and the local contrast loss of the plurality of news in the second field in the test set are similar to the method steps of determining the global contrast loss and the local contrast loss of the plurality of news in the training set, and will not be described in detail in the embodiments of the present application.
Further, after obtaining global contrast loss and local contrast loss of the plurality of news in the second domain in the test set, determining a difference between the news spread map features in the second domain in the training set and the news spread map features in the second domain in the test set based on the data adaptive constraint module in the social media false news detection model; and then, the difference between the news propagation graph characteristics of the second field in the training set and the news propagation graph characteristics of the second field in the test set, and the weighted sum of the global contrast loss and the local contrast loss of a plurality of news in the second field in the test set are used as the value of a second target loss function, and the social media false news detection model is trained based on the value of the second target loss function, so that the social media false news detection model based on the potential characteristics of the test set is obtained. The feature extractor on the test set and the feature extractor on the training set are similar in performance, optimization of the feature extractor is achieved, the occurrence of the fitting problem is prevented, and therefore the false news in the second field to be detected can be accurately identified by adding the social media false news detection model after the constraint on the difference between the news propagation diagram features in the training set and the news propagation diagram features in the test set. Optionally, if the constraint of the difference between the news propagation graph feature of the second domain in the training set and the news propagation graph feature of the second domain in the test set is not added in the loss function, the overfitting problem may occur, so that the social media false news detection model trained based on the training set cannot accurately detect the authenticity of the second domain news to be detected, and thus the false news of the second domain cannot be accurately and effectively identified.
According to the method, constraint of difference between the news spread graph characteristics of the second field in the training set and the news spread graph characteristics of the second field in the test set is added in the loss function, so that fine adjustment of the social media false news detection model and optimization of the feature extractor are achieved, the feature extractor on the test set and the feature extractor on the training set are similar in performance, the occurrence of the fitting problem is effectively prevented, and therefore the social media false news detection model after constraint is conducted on the difference between the news spread graph characteristics in the training set and the news spread graph characteristics in the test set is added can more accurately identify false news in the second field to be detected.
The embodiment of the application provides a social media false news detection method, which comprises the following steps:
acquiring a news propagation diagram of a second field to be detected;
inputting the news propagation diagram of the second field to be detected into a social media false news detection model to obtain an authenticity detection result of social media news corresponding to the news propagation diagram of the second field; the social media false news detection model is trained based on the training method of the social media false news detection model.
Specifically, after training the social media false news detection model based on the training set and the test set, the detection of the second-domain false news can be accurately performed based on the trained social media false news detection model. Optionally, the news propagation diagram of the second field to be detected can be firstly obtained, then the news propagation diagram of the second field to be detected is input into the trained social media false news detection model, and the authenticity detection result of the social media news corresponding to the news propagation diagram of the second field can be obtained, so that the authenticity detection of the second field news with less data quantity and shorter propagation time is realized, and the accuracy of the second field false news detection with less data quantity and shorter propagation time is improved.
Exemplary, a specific flow of a training method of the social media false news detection model in the embodiment of the present application is shown in fig. 3 and fig. 4, and specifically is as follows:
1. training of the social media false news detection model is performed based on the first domain (high resource) news in the training set.
By performing three different types of augmentation operations on the first-domain news data: edge loss, feature disorder and feature masking to generate three different types of augmentation graphs. A feature extractor is used to obtain a high-dimensional feature matrix of the first domain news data and the three types of augmentation charts. The primary task and the secondary task share a feature extractor. The method comprises the steps of inputting first-field false news data into a main task, inputting the first-field news data and three types of augmentation graphs into an auxiliary task, performing local contrast learning on the augmentation graphs subjected to edge losing operation and feature random masking operation, and performing global contrast learning on the first-field news data and the augmentation graphs subjected to feature disturbing operation.
1.1 Global contrast learning based on a graph roll-up neural network
The purpose of global contrast learning is to help nodes in the graph data represent global information that gets the entire graph. Given an input eventBy each ofThe type of data augmentation may produce different graphs. For global contrast learning, two views are employed: one is original View0; the other is the enhanced View1, where node attributes of all nodes in the graph are randomly assigned. With both graphs, two corresponding node representations Z0 and Z1 can be obtained by a shared feature extractor. Thereafter, a global graph representation s can be summarized from the node representation matrix Z0 of the original View View0 by the multi-layer perceptron. />
The positive samples in global contrast learning consist of node-graph representation pairs, where both the node representation and the graph representation are from the original View0. The negative samples are also composed of node-graph representation pairs, where the node representation is from View1 and the graph representation is from View0. The arbiter D calculates the scores of the positive and negative samples, respectively, which should be higher for positive samples and lower for negative samples. Wherein the representation of D is defined as follows:
D(Z si ,s)＝Sigmoid(Z si *s)
Wherein Z is si Node representation representing Views, represents inner product. The loss function for global contrast learning is as follows:
where N represents the number of nodes in the input graph.
1.2 local contrast learning based on a graph roll-up neural network
Through local contrast learning, the model can learn a richer node characteristic representation. By comparing the characteristic difference between the nodes and the neighbor nodes, the model can learn the node representation with more distinguishing property, can reduce the influence of the node position, noise, missing value and the like on the graph data, and improves the robustness and the interpretability of the graph data. Through local contrast learning, the model can better capture the similarity and the difference between the nodes and the neighbor nodes, so that the context information of the nodes can be better understood.
Given an input eventTwo different augmentation graphs are obtained as inputs to the feature extractor by graph augmentation policies (edge deletion and node attribute masking). The output of the feature extractor is two node representation matrices, Z 2 ，Z 3 . The basic goal of local contrast learning is to distinguish whether two nodes from the enhanced view are the same node, and therefore, (Z) 2i ,Z 3i ) (i e {1,., N }) represents the dead-end, where N is the number of nodes, (Z 2i ,Z 3j ) And (Z) 2i ,Z 2j ) (i, j e {1,., N }, i+.j) represents intra-class negative pairs and inter-class negative pairs. The facing objective function is defined as follows:
where cos () represents the cosine similarity function, τ is a hyper-argument, and g () is a two-layer MLP. The nodes passing through the MLP are denoted as Z2 'and Z3', and regularization modifiers are used on the redefined node representation. The following are provided:
the local contrast final loss function is defined as follows:
the training method based on the high-resource false news data trained during the test is characterized in that the final loss function of the auxiliary task is as follows:
L s ＝L g +αL l
the training method based on the high-resource false news data trained during the test has the following loss function:
L＝L m +γL s
where Lm is the loss of the primary task.
2. Training of the social media false news detection model is performed based on the second domain (low resource) news in the training set.
Three different types of augmentation operations are performed on the second-domain news data: edge loss, feature disorder and feature masking to generate three different types of augmentation graphs. A feature extractor is used to obtain a high-dimensional feature matrix of the second-domain news data and the three types of augmentation charts. The feature extractor here is the same as mentioned in the first section. And inputting the second-field news data and the three types of augmentation graphs into an auxiliary task, performing local contrast learning on the augmentation graphs subjected to edge losing operation and feature random masking operation, and performing global contrast learning on the second-field false news data and the augmentation graphs subjected to feature scrambling operation. The training of the social media false news detection model based on the news in the second field in the training set also adopts a self-supervision learning framework comprising global contrast learning and local contrast learning, and the self-supervision learning framework is the same as the model of the auxiliary task in the training method for the social media false news detection model based on the news in the first field in the training set, and the auxiliary task is not repeated here.
3. A data self-adaptive constraint method based on training at test time.
And obtaining a feature matrix of the news data by using READOUT function, and calculating two statistical data, a feature average value and a covariance matrix of the feature matrix of the news data. And calculating two statistical data, eigenvalue and covariance matrix of the false news data in the second field one by one, and constructing a loss function.
Definition of the definitionUsing READOUT functions (READOUT functions are feature extractors)
Obtaining a feature matrix of the news data, calculating two statistical data of the feature matrix of the news data, wherein the definition of a feature average value and a covariance matrix is as follows:
calculating a feature matrix of low-resource false news data in the second field through READOUT function, and also calculating two statistical data, a feature average value and a covariance matrix, and constructing a loss function as follows:
wherein mu t ,∑ t The characteristic average and covariance matrices of the low-resource false news data are represented, respectively.
According to the method, whether the news data is false news is judged by acquiring the news data to be detected and inputting the news data into the false news detection model. According to the model training method, the characteristics of the false news data in two different fields are fused into the model representation, so that better modeling of the target data field is realized, the target data is trained again in the test stage, the representation capacity of the model on the target data is enhanced, the false news can be detected more accurately and efficiently, and the accuracy of cross-domain detection of the false news is improved.
The training device of the social media false news detection model provided by the invention is described below, and the training device of the social media false news detection model described below and the training method of the social media false news detection model described above can be correspondingly referred to each other.
Fig. 5 is a schematic structural diagram of a training device of a social media false news detection model provided by the invention. The training device of the social media false news detection model provided in this embodiment includes:
a first obtaining module 710, configured to obtain a plurality of news propagation maps of a first domain in a training set; the news propagation graph is used for representing a news propagation path;
a second obtaining module 720, configured to obtain a plurality of news propagation maps in a second field in the training set; the ratio of the number of news maps of the second domain to the number of news maps of the first domain is less than a threshold;
the training module 730 is configured to train the social media false news detection model according to the plurality of news propagation graphs in the first domain, the plurality of news propagation graphs in the second domain, and the first objective loss function, to obtain a social media false news detection model trained based on the training set; the social media false news detection model is used for detecting the authenticity of news; the first objective loss function is determined from the classification loss, the global contrast loss, and the local contrast loss; the global contrast loss represents the node characteristics in the news propagation graph, and the degree of association between the node characteristics in the first type of the news propagation graph and the news propagation graph characteristics; the local contrast loss represents a degree of association between node features in a second type of augmentation graph of the news propagation graph and node features in a third type of augmentation graph of the news propagation graph; the classification loss indicates the accuracy of the classification result.
Optionally, the social media false news detection model includes at least one of the following:
a feature extraction module; the feature extraction module is used for extracting news propagation graph features;
a classification module; the classification module is used for predicting the authenticity of social media news corresponding to the news propagation graph according to the news propagation graph characteristics;
a self-supervision learning module; the self-supervision learning module is used for determining global contrast loss according to the node characteristics in the news propagation graph, the node characteristics in the first type of the news propagation graph and the news propagation graph characteristics; and determining local contrast loss according to the node characteristics in the second type augmentation chart of the news propagation chart and the node characteristics in the third type augmentation chart of the news propagation chart.
Optionally, the social media false news detection model is trained based on the following:
inputting a plurality of news propagation graphs of a first field and a plurality of news propagation graphs of a second field in a training set into a social media false news detection model, and outputting an authenticity detection result of social media news corresponding to the news propagation graphs; obtaining classification loss of the social media false news detection model according to the authenticity detection result of the social media news and the label information of the news; the tag information is used for marking the authenticity of news;
Inputting a plurality of news propagation graphs in the first field, a first type of augmentation graph corresponding to the plurality of news propagation graphs in the first field, a second type of augmentation graph corresponding to the plurality of news propagation graphs in the first field and a third type of augmentation graph corresponding to the plurality of news propagation graphs in the first field into a social media false news detection model to obtain global contrast loss and local contrast loss of a plurality of news in the first field;
inputting a plurality of news propagation graphs in the second field in the training set, a first type augmentation graph corresponding to the plurality of news propagation graphs in the second field, a second type augmentation graph corresponding to the plurality of news propagation graphs in the second field and a third type augmentation graph corresponding to the plurality of news propagation graphs in the second field into a social media false news detection model to obtain global contrast loss and local contrast loss of a plurality of news in the second field;
taking the weighted sum of the classification loss of the social media false news detection model, the global contrast loss and the local contrast loss of the news in the first field, and the global contrast loss and the local contrast loss of the news in the second field as the value of a first target loss function;
Training the social media false news detection model based on the value of the first target loss function to obtain the social media false news detection model trained based on the training set.
Optionally, the social media false news detection model further includes:
a data self-adaptive constraint module; the data adaptive constraint module is used for determining differences between the news spread map features in the training set and the news spread map features in the test set.
Optionally, the training module 730 is further configured to input a plurality of news propagation graphs in the second field in the test set, a first type of augmentation graph corresponding to the plurality of news propagation graphs in the second field in the test set, a second type of augmentation graph corresponding to the plurality of news propagation graphs in the second field in the test set, and a third type of augmentation graph corresponding to the plurality of news propagation graphs in the second field in the test set to a social media false news detection model trained based on the training set, so as to obtain global contrast loss and local contrast loss of the plurality of news in the second field in the test set;
taking the sum of the global contrast loss and the local contrast loss weights of a plurality of news in the second field in the test set as a value of a second target loss function according to the difference between the news propagation map features in the second field in the training set and the news propagation map features in the second field in the test set;
Training the social media false news detection model based on the value of the second target loss function to obtain the social media false news detection model based on the potential characteristics of the test set.
The device of the embodiment of the present invention is configured to perform the method of any of the foregoing method embodiments, and its implementation principle and technical effects are similar, and are not described in detail herein.
Fig. 6 illustrates a physical schematic diagram of an electronic device, which may include: processor 810, communication interface (Communications Interface) 820, memory 830, and communication bus 840, wherein processor 810, communication interface 820, memory 830 accomplish communication with each other through communication bus 840. The processor 810 may invoke logic instructions in the memory 830 to perform a training method for a social media false news detection model, the method comprising: acquiring a plurality of news propagation graphs of a first field in a training set; the news propagation graph is used for representing a news propagation path; acquiring a plurality of news propagation graphs of a second field in the training set; the ratio of the number of news maps of the second domain to the number of news maps of the first domain is less than a threshold; training the social media false news detection model according to the plurality of news propagation graphs in the first field, the plurality of news propagation graphs in the second field and the first target loss function to obtain a social media false news detection model trained based on the training set; the social media false news detection model is used for detecting the authenticity of news; the first objective loss function is determined from the classification loss, the global contrast loss, and the local contrast loss; the global contrast loss represents the node characteristics in the news propagation graph, and the degree of association between the node characteristics in the first type of the news propagation graph and the news propagation graph characteristics; the local contrast loss represents a degree of association between node features in a second type of augmentation graph of the news propagation graph and node features in a third type of augmentation graph of the news propagation graph; the classification loss indicates the accuracy of the classification result.
Further, the logic instructions in the memory 830 described above may be implemented in the form of software functional units and may be stored in a computer-readable storage medium when sold or used as a stand-alone product. Based on this understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.
In another aspect, the present invention also provides a computer program product comprising a computer program stored on a non-transitory computer readable storage medium, the computer program comprising program instructions which, when executed by a computer, enable the computer to perform a training method of a social media false news detection model provided by the methods described above, the method comprising: acquiring a plurality of news propagation graphs of a first field in a training set; the news propagation graph is used for representing a news propagation path; acquiring a plurality of news propagation graphs of a second field in the training set; the ratio of the number of news maps of the second domain to the number of news maps of the first domain is less than a threshold; training the social media false news detection model according to the plurality of news propagation graphs in the first field, the plurality of news propagation graphs in the second field and the first target loss function to obtain a social media false news detection model trained based on the training set; the social media false news detection model is used for detecting the authenticity of news; the first objective loss function is determined from the classification loss, the global contrast loss, and the local contrast loss; the global contrast loss represents the node characteristics in the news propagation graph, and the degree of association between the node characteristics in the first type of the news propagation graph and the news propagation graph characteristics; the local contrast loss represents a degree of association between node features in a second type of augmentation graph of the news propagation graph and node features in a third type of augmentation graph of the news propagation graph; the classification loss indicates the degree of correctness of the classification result.
In yet another aspect, the present invention further provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, is implemented to perform the above-described methods of training the social media false news detection model.
The apparatus embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art will understand and implement the present invention without undue burden.
From the above description of the embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented by means of software plus necessary general hardware platforms, or of course may be implemented by means of hardware. Based on this understanding, the foregoing technical solution may be embodied essentially or in a part contributing to the prior art in the form of a software product, which may be stored in a computer readable storage medium, such as ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the method described in the respective embodiments or some parts of the embodiments.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. The training method of the social media false news detection model is characterized by comprising the following steps of:
acquiring a plurality of news propagation graphs of a first field in a training set; the news propagation graph is used for representing a news propagation path;
acquiring a plurality of news propagation graphs of a second field in the training set; the ratio of the number of the news broadcasting pictures in the second field to the number of the news broadcasting pictures in the first field is smaller than a threshold value;
training the social media false news detection model according to the plurality of news propagation graphs in the first field, the plurality of news propagation graphs in the second field and the first target loss function to obtain a social media false news detection model trained based on a training set; the social media false news detection model is used for detecting the authenticity of news; the first objective loss function is determined from classification loss, global contrast loss, and local contrast loss; the global contrast loss represents the node characteristics in the news propagation graph, and the degree of association between the node characteristics in the first type of the news propagation graph and the news propagation graph characteristics; the local contrast loss represents a degree of association between node features in a second type of augmentation graph of the news propagation graph and node features in a third type of augmentation graph of the news propagation graph; the classification loss represents the accuracy of the classification result.
2. The method of training a social media false news detection model according to claim 1, wherein the social media false news detection model comprises at least one of:
a feature extraction module; the feature extraction module is used for extracting news propagation graph features;
a classification module; the classification module is used for predicting the authenticity of social media news corresponding to the news propagation graph according to the news propagation graph characteristics;
a self-supervision learning module; the self-supervision learning module is used for determining the global contrast loss according to the node characteristics in the news propagation graph, the node characteristics in the first type of the news propagation graph and the news propagation graph characteristics; and determining the local contrast loss according to the node characteristics in the second type augmentation chart of the news propagation chart and the node characteristics in the third type augmentation chart of the news propagation chart.
3. The training method of a social media false news detection model according to claim 2, wherein the social media false news detection model is trained based on the following manner:
inputting a plurality of news propagation graphs of a first field and a plurality of news propagation graphs of a second field in a training set into a social media false news detection model, and outputting an authenticity detection result of social media news corresponding to the news propagation graphs; obtaining the classification loss of the social media false news detection model according to the authenticity detection result of the social media news and the label information of the news; the tag information is used for marking the authenticity of the news;
Inputting a plurality of news propagation graphs of the first field, a first type of augmentation graph corresponding to the plurality of news propagation graphs of the first field, a second type of augmentation graph corresponding to the plurality of news propagation graphs of the first field and a third type of augmentation graph corresponding to the plurality of news propagation graphs of the first field in a training set to a social media false news detection model to obtain global contrast loss and local contrast loss of a plurality of news of the first field;
inputting a plurality of news propagation graphs of the second field, a first type of augmentation graph corresponding to the plurality of news propagation graphs of the second field, a second type of augmentation graph corresponding to the plurality of news propagation graphs of the second field and a third type of augmentation graph corresponding to the plurality of news propagation graphs of the second field in a training set to a social media false news detection model to obtain global contrast loss and local contrast loss of a plurality of news of the second field;
taking the weighted sum of the classification loss of the social media false news detection model, the global contrast loss and the local contrast loss of a plurality of news in the first field, and the global contrast loss and the local contrast loss of a plurality of news in the second field as the value of a first target loss function;
And training the social media false news detection model based on the value of the first target loss function to obtain a social media false news detection model trained based on a training set.
4. The method for training a social media false news detection model according to claim 3, further comprising:
a data self-adaptive constraint module; the data adaptive constraint module is used for determining differences between the news spread map features in the training set and the news spread map features in the test set.
5. The method for training a social media false news detection model according to claim 4, wherein training the social media false news detection model based on the value of the first objective loss function, after obtaining a social media false news detection model trained based on a training set, further comprises:
inputting a plurality of news propagation graphs in a second field in a test set, a first type augmentation graph corresponding to the plurality of news propagation graphs in the second field in the test set, a second type augmentation graph corresponding to the plurality of news propagation graphs in the second field in the test set and a third type augmentation graph corresponding to the plurality of news propagation graphs in the second field in the test set into the social media false news detection model trained based on the training set to obtain global contrast loss and local contrast loss of a plurality of news in the second field in the test set;
Taking the sum of the global contrast loss and the local contrast loss weights of a plurality of news in the second field in the test set as a value of a second target loss function according to the difference between the news propagation map features in the second field in the training set and the news propagation map features in the second field in the test set;
and training the social media false news detection model based on the value of the second target loss function to obtain the social media false news detection model based on the potential characteristics of the test set.
6. A method for detecting false news of social media, comprising:
acquiring a news propagation diagram of a second field to be detected;
inputting the news propagation diagram of the second field to be detected into the social media false news detection model to obtain an authenticity detection result of social media news corresponding to the news propagation diagram of the second field; the social media false news detection model is trained based on the method of any one of claims 1-5.
7. A training device for a social media false news detection model, comprising:
the first acquisition module is used for acquiring a plurality of news propagation graphs of a first field in the training set; the news propagation graph is used for representing a news propagation path;
The second acquisition module is used for acquiring a plurality of news propagation graphs in a second field in the training set; the ratio of the number of the news broadcasting pictures in the second field to the number of the news broadcasting pictures in the first field is smaller than a threshold value;
the training module is used for training the social media false news detection model according to the plurality of news propagation graphs in the first field, the plurality of news propagation graphs in the second field and the first target loss function to obtain a social media false news detection model trained based on a training set; the social media false news detection model is used for detecting the authenticity of news; the first objective loss function is determined from classification loss, global contrast loss, and local contrast loss; the global contrast loss represents the node characteristics in the news propagation graph, and the degree of association between the node characteristics in the first type of the news propagation graph and the news propagation graph characteristics; the local contrast loss represents a degree of association between node features in a second type of augmentation graph of the news propagation graph and node features in a third type of augmentation graph of the news propagation graph; the classification loss represents the accuracy of the classification result.
8. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the training method of the social media false news detection model of any one of claims 1 to 5 or the social media false news detection method of claim 6 when the program is executed by the processor.
9. A non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor implements the training method of the social media false news detection model according to any one of claims 1 to 5 or the social media false news detection method according to claim 6.
10. A computer program product having stored thereon executable instructions that, when executed by a processor, cause the processor to implement the training method of a social media false news detection model according to any one of claims 1 to 5 or the social media false news detection method according to claim 6.
CN202311475479.XA 2023-11-07 2023-11-07 Training method of social media false news detection model Pending CN117576504A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311475479.XA CN117576504A (en) 2023-11-07 2023-11-07 Training method of social media false news detection model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311475479.XA CN117576504A (en) 2023-11-07 2023-11-07 Training method of social media false news detection model

Publications (1)

Publication Number Publication Date
CN117576504A true CN117576504A (en) 2024-02-20

Family

ID=89861585

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311475479.XA Pending CN117576504A (en) 2023-11-07 2023-11-07 Training method of social media false news detection model

Country Status (1)

Country Link
CN (1) CN117576504A (en)

Similar Documents

Publication Publication Date Title
CN111582409B (en) Training method of image tag classification network, image tag classification method and device
Thieltges et al. The devil’s triangle: Ethical considerations on developing bot detection methods
CN113870254B (en) Target object detection method and device, electronic equipment and storage medium
TWI803243B (en) Method for expanding images, computer device and storage medium
CN116311214B (en) License plate recognition method and device
CN111144425B (en) Method and device for detecting shot screen picture, electronic equipment and storage medium
CN111259823A (en) Pornographic image identification method based on convolutional neural network
CN109101984B (en) Image identification method and device based on convolutional neural network
CN117409419A (en) Image detection method, device and storage medium
CN111915595A (en) Image quality evaluation method, and training method and device of image quality evaluation model
CN113723352B (en) Text detection method, system, storage medium and electronic equipment
CN117558011B (en) Image text tampering detection method based on self-consistency matrix and multi-scale loss
CN114821823A (en) Image processing, training of human face anti-counterfeiting model and living body detection method and device
CN111950629A (en) Method, device and equipment for detecting confrontation sample
CN112613341A (en) Training method and device, fingerprint identification method and device, and electronic device
CN111967383A (en) Age estimation method, and training method and device of age estimation model
CN116433909A (en) Similarity weighted multi-teacher network model-based semi-supervised image semantic segmentation method
CN116232699A (en) Training method of fine-grained network intrusion detection model and network intrusion detection method
CN117576504A (en) Training method of social media false news detection model
CN111209567B (en) Method and device for judging perceptibility of improving robustness of detection model
CN113610080A (en) Cross-modal perception-based sensitive image identification method, device, equipment and medium
Wang et al. Saliency detection by multilevel deep pyramid model
Nam et al. Spam Image Detection Model based on Deep Learning for Improving Spam Filter
CN117830728B (en) Abnormal flow detection method, device, electronic equipment and medium
CN113205082B (en) Robust iris identification method based on acquisition uncertainty decoupling

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination