A multi-view collaborative filtering method for detecting fake news based on user post relationships.

Abstract

This invention discloses a multi-view collaborative filtering method for detecting fake news based on user post relationships, belonging to the field of fake news detection technology. The key technical points are: utilizing the difficult-to-fake user-post relationship to model user preferences for news authenticity; specifically, a 3D matrix is ​​constructed to facilitate user-to-post and post-to-user propagation within each individual view; and multiple views are constructed for multiple post tags, while contrastive learning is employed in the global view. Furthermore, this invention provides two strategies for integrating multiple views to comprehensively capture advanced user behavior features: the first strategy utilizes meta-learning to obtain appropriate user preference weights, and the second strategy utilizes standard deviation metrics to evaluate the explicitness of user preferences; finally, fake news is identified using post features and user behavior features, fully leveraging different user preference features to improve the efficiency of news authenticity feature extraction and increase the accuracy of fake news identification.

Description

Technical Field

[0001] This invention relates to the field of fake news detection technology, and more specifically, to a multi-view collaborative filtering method for detecting fake news based on user post relationships. Background Technology

[0002] In recent years, with the development of social media, more and more people are publishing or browsing various news through online platforms, which helps the public to obtain the latest news information in a timely manner. However, at the same time, some people publish fake news for profit, which reduces the security and reliability of news dissemination and affects user experience.

[0003] Existing fake news detection methods based on user modeling mostly rely on static attributes or historical posting behavior of users who post or participate in the dissemination of posts to determine users' preferences for the authenticity of news. However, in practice, these static attributes and historical posting behavior are very easy to fake. Some malicious users may use this information to deceive the model, greatly reducing the efficiency of fake news detection models.

[0004] A small number of fake news detection methods based on user modeling attempt to model users' social networks to mine the interaction features between users, because users' social networks have great potential for judging the theme of events. However, in the field of fake news detection, users' social networks introduce a lot of noise, which leads to inaccurate extraction and detection of the authenticity features of news. Summary of the Invention

[0005] The purpose of this invention is to solve the above-mentioned problems and provide a multi-view collaborative filtering method for detecting fake news based on user post relationships. This method utilizes user-post relationships that are difficult to fake to model user preferences for the authenticity of news. Specifically, a 3D matrix is ​​constructed to facilitate user-to-post and post-to-user propagation in each individual view; and multiple views are constructed for multiple post tags, while contrastive learning is employed in the global view.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution: a multi-view collaborative filtering method for detecting fake news based on user post relationships, the method comprising the following steps:

[0007] S1: Process the source text of the news to obtain its textual features;

[0008] S2: Construct multiple views based on the propagation relationship between users and posts in posts with different tags, and obtain the user and post representation of the overall view by passing information in a single view;

[0009] S3: Calculate the standard deviation of user representations among the sub-views as the weight of user preference explicitness to aggregate user representations in the overall view to obtain the explicit representation of user preferences;

[0010] S4: Use meta-learning to train and obtain user preference weights Wp to aggregate the user representations of each subview to obtain the user preference representation of each user. For a post, take the average of all user preference representations participating in the post as the user preference representation of the post.

[0011] S5: Concatenate all the features from S1 to S4 to obtain the final features of the post, and design and implement a fully connected neural network to predict the true or false tags of the news.

[0012] S6: Train the entire network obtained in S5 using cross-entropy loss.

[0013] The present invention is further configured such that S1 is achieved through the following steps:

[0014] S1-1: Transfer the source text data {T1,T2,…,T...} m} is converted into its corresponding Word2Vec vector representation, denoted as W1, W2, ..., W m};

[0015] S1-2: Use convolutional neural networks (CNNs) with kernel sizes of {3, 4, 5} to process {W1, W2, ..., W... m The text is processed and average pooling is performed to obtain the corresponding text features {t1,t2,…,t} of the source text. m}

[0016] The present invention is further configured such that S2 is implemented through the following steps;

[0017] S2-1: Construct multiple views; use a three-dimensional matrix to represent the user-post relationship under different tags, such as [i,j,k] indicating that the i-th user participated in the j-th post and the post's tag is the k-th category; thus, we can obtain a total view and multiple sub-views; the total view does not classify post tags and only considers the user-post relationship, while the sub-views correspond to the user-post relationship under different tags respectively;

[0018] S2-2: Performing message passing operations in a single view yields the user representation and post representation for the next layer, using the following formula:

[0019]

[0020]

[0021] S2-3: Perform two message passing operations in a single view, and average the user representation and post representation of each layer to obtain the user representation and post representation of the single view;

[0022] S2-4: After the operations described in S2-1 to S2-3, the user representation of the overall view {U} is obtained. G 1, U G 2, ..., U G n}, the post represents {P} G 1, P G 2, ..., P G m} and the user representation of each subview {U Si 1, U Si 2, ..., U Si n}, i represents the i-th subview, and the post is represented as {P}. G 1, P G 2, ..., P G m} is a feature of the post {p1,p2,…,p} m}

[0023] The present invention is further configured such that S3 is implemented through the following steps;

[0024] S3-1: Calculate the standard deviation W of the user representation among the subviews obtained in S2. s As a weight for the explicitness of user preferences;

[0025] S3-2: Utilizing the weight W for the explicitness of user preferences s The aggregated view of users yields a specific representation of user preferences for each post, {s1, s2, ..., s}. m}

[0026] The present invention is further configured such that S4 is implemented through the following steps;

[0027] S4-1: Utilizing meta-learning to initialize user preference weights W for users with more user-post behavior. p And migrate it to users with less posting activity;

[0028] S4-2: Utilizing user preference weights W p Aggregating the user representations of each subview yields a user preference representation {U} R 1, U R 2, ..., U R n};

[0029] S4-3: For any post, take the mean of all user preference representations participating in that post as the user preference representation {r1, r2, ..., r} for that post. m}

[0030] In summary, the present invention has the following beneficial effects:

[0031] 1. This invention is based on user-post behavior and uses a collaborative filtering algorithm to model user propagation behavior in posts with different tags. From the perspective of user dynamic behavior, it solves the problem that user static attributes and historical posting behavior are easily disguised, which leads to inaccurate user behavior modeling. This enables the model to better capture user behavior characteristics and achieve a higher fake news detection rate.

[0032] 2. This invention utilizes the user-post relationship, which is difficult to fake, to model users' preferences for the authenticity of news. Specifically, a 3D matrix is ​​constructed to facilitate user-to-post and post-to-user propagation in each single view; and multiple views are constructed for multiple post tags, while contrastive learning is employed in the global view.

[0033] 3. This invention provides two strategies for integrating multiple views to comprehensively capture advanced user behavior features. The first strategy utilizes meta-learning to obtain appropriate user preference weights, and the second strategy utilizes a standard deviation metric to evaluate the explicitness of user preferences. Finally, the learned post features and user behavior features are used for fake news identification, fully leveraging different user preference features, thus improving the efficiency of extracting the authenticity features of news and increasing the accuracy of fake news identification. Attached Figure Description

[0034] Figure 1 This is a structural diagram of the fake news detection model constructed in this embodiment of the invention. Detailed Implementation

[0035] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be described in further detail below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0036] It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of this invention can be combined with each other. The invention will now be described in detail with reference to the embodiments.

[0037] Example:

[0038] like Figure 1 As shown, a multi-view collaborative filtering method for detecting fake news based on user post relationships includes the following steps:

[0039] S1: Use CNN to process the source text of the news to obtain its text features;

[0040] S1-1: Transfer the source text data {T1,T2,…,T...} m} is converted into its corresponding Word2Vec vector representation, denoted as W1, W2, ..., W m};

[0041] S1-2: Use convolutional neural networks (CNNs) with kernel sizes of {3, 4, 5} to process {W1, W2, ..., W... m The text is processed and average pooling is performed to obtain the corresponding text features {t1,t2,…,t} of the source text. m}

[0042] S2: Construct multiple views based on the propagation relationship between users and posts in posts with different tags, and obtain the user and post representation of the overall view by passing information in a single view;

[0043] S2-1: Construct multiple views; use a three-dimensional matrix to represent the user-post relationship under different tags, such as [i,j,k] indicating that the i-th user participated in the j-th post and the post's tag is the k-th category; thus, we can obtain a total view and multiple sub-views; the total view does not classify post tags and only considers the user-post relationship, while the sub-views correspond to the user-post relationship under different tags respectively;

[0044] S2-2: Performing message passing operations in a single view yields the user representation and post representation for the next layer, using the following formula:

[0045]

[0046]

[0047] S2-3: Perform two message passing operations in a single view, and average the user representation and post representation of each layer to obtain the user representation and post representation of the single view;

[0048] S2-4: After operations S2-1 to S2-3, the user representation of the overall view {U} is obtained. G 1, U G 2, ..., U G n}, the post represents {P} G 1, P G 2, ..., P G m} and the user representation of each subview {U Si 1, U Si 2, ..., U Si n}, i represents the i-th subview; and the post is represented as {P} G1, P G 2, ..., P G m} is a feature of the post {p1,p2,…,p} m}

[0049] S3: Calculate the standard deviation of user representations among the sub-views as the weight of user preference explicitness to aggregate user representations in the overall view to obtain the explicit representation of user preferences;

[0050] S3-1: Calculate the standard deviation W of the user representation among the subviews obtained in S2. s As a weight for the explicitness of user preferences;

[0051] S3-2: Utilizing the weight W for the explicitness of user preferences s The aggregated view of users yields a specific representation of user preferences for each post, {s1, s2, ..., s}. m}

[0052] S4: Use meta-learning to train and obtain user preference weights Wp to aggregate the user representations of each subview to obtain the user preference representation of each user. For a post, take the average of all user preference representations participating in the post as the user preference representation of the post.

[0053] S4: This is achieved through the following steps;

[0054] S4-1: Utilizing meta-learning to initialize user preference weights W for users with more user-post behavior. p And migrate it to users with less posting activity;

[0055] S4-2: Utilizing user preference weights W p Aggregating the user representations of each subview yields a user preference representation {U} R 1, U R 2, ..., U R n};

[0056] S4-3: For any post, take the mean of all user preference representations participating in that post as the user preference representation {r1, r2, ..., r} for that post. m}

[0057] S5: Concatenate all the features from S1 to S4 to obtain the final features of the post. Design and implement a fully connected neural network to predict the true or false tags of the news.

[0058] S6: Train the entire network obtained from S5 using cross-entropy loss.

[0059] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A multi-view collaborative filtering method for detecting fake news based on user post relationships, characterized by: The method includes the following steps: S1: Process the source text of the news to obtain its textual features; S2: Construct multiple views based on the propagation relationship between users and posts in posts with different tags, and obtain the user and post representation of the overall view by passing information in a single view; S2 is achieved through the following steps; S2-1: Construct multiple views; use a three-dimensional matrix to represent the user-post relationship under different tags, such as [i,j,k] indicating that the i-th user participated in the j-th post and the post's tag is the k-th category; thus, we can obtain a total view and multiple sub-views; the total view does not classify post tags and only considers the user-post relationship, while the sub-views correspond to the user-post relationship under different tags respectively; S2-2: Performing message passing operations in a single view yields the user representation and post representation for the next layer, using the following formula: S2-3: Perform two message passing operations in a single view, and average the user representation and post representation of each layer to obtain the user representation and post representation of the single view; S2-4: After the operations described in S2-1 to S2-3, the user representation of the overall view {U} is obtained. G 1, U G 2, ..., U G n}, the post represents {P} G 1, P G 2, ..., P G m} and the user representation of each subview {U Si 1, U Si 2, ..., U Si n}, i represents the i-th subview; and the post is represented as {P}. G 1, P G 2, ..., P G m} is a feature of the post {p1, p2, …, p} m }; S3: Calculate the standard deviation of user representations among the sub-views as the weight of user preference explicitness to aggregate user representations in the overall view to obtain the explicit representation of user preferences; S4: Use meta-learning to train and obtain user preference weights Wp to aggregate the user representations of each subview to obtain the user preference representation of each user. For a post, take the average of all user preference representations participating in the post as the user preference representation of the post. S4 is achieved through the following steps; S4-1: Utilizing meta-learning to initialize user preference weights W for users with more user-post behavior. p And migrate it to users with less posting activity; S4-2: Utilizing user preference weights W p Aggregating the user representations of each subview yields a user preference representation {U} R 1, U R 2, ..., U R n}; S4-3: For any post, take the mean of all user preference representations participating in that post as the user preference representation {r1, r2, …, r} for that post. m }; S5: Concatenate all the features from S1 to S4 to obtain the final features of the post, and design and implement a fully connected neural network to predict the true or false tags of the news. S6: Train the entire network obtained in S5 using cross-entropy loss.

2. The method for detecting fake news based on multi-view collaborative filtering of user post relationships according to claim 1, characterized in that: S1 is achieved through the following steps: S1-1: Transfer the source text data {T1, T2, …, T} m } is converted into its corresponding Word2Vec vector representation, denoted as W1, W2, …, W m }; S1-2: Use convolutional neural networks (CNNs) with kernel sizes of {3, 4, 5} to process {W1, W2, …, W m The data is processed and average pooling is performed to obtain the corresponding text features {t1, t2, …, t} of the source text. m } 3. The method for detecting fake news based on multi-view collaborative filtering of user post relationships according to claim 1, characterized in that: S3 is achieved through the following steps; S3-1: Calculate the standard deviation W of the user representation among the subviews obtained in S2. s As a weight for the explicitness of user preferences; S3-2: Utilizing the weight W for the explicitness of user preferences s The aggregated view of users yields a specific representation of user preferences for each post, {s1, s2, …, s}. m }

Images