A Method and System for Detecting Fake News Based on Time-Series Trend Prediction

Abstract

This invention proposes a method for detecting fake news based on time-series trend prediction, comprising: extracting news representations from news samples; clustering the news representations to obtain at least one topic cluster; modeling the news samples of each topic cluster according to their time-series distribution into a data model corresponding to the news topic; predicting the time-series distribution trend of the news topic based on the data model; reweighting the news samples of each topic cluster based on the time-series distribution trend and merging them into a training set; training a fake news detector using the training set; and detecting fake news in target news using the fake news detector. This invention also proposes a fake news detection system based on time-series trend prediction, and a data processing device for implementing fake news detection based on time-series trend prediction.

Description

Technical Field

[0001] This invention belongs to the field of cyberspace security technology, and in particular relates to a method and system for detecting fake news based on time-series trend prediction. Background Technology

[0002] Automated fake news detection aims to automatically identify news containing inaccurate and deliberately misleading information from a pool of news articles, a crucial task in maintaining a healthy online news ecosystem. As a supplement to human verification, automated fake news detection can efficiently filter fake news from a vast news database.

[0003] For the past decade, most researchers in fake news detection have followed a traditional paradigm: collecting a fixed dataset and randomly dividing it into training and test sets. However, in practice, fake news detection models can only be trained on data collected up to the current time (i.e., offline data) to detect fake news in the next time period (i.e., online data). Due to rapid changes in the external environment, the distribution of news topics often changes over time, leading to distribution discrepancies between offline and online data—a phenomenon known as temporal drift. Recent empirical studies have shown that the performance of fake news detection models significantly degrades when the dataset is split from random to time-ordered. Therefore, the temporal drift problem has been a key obstacle to applying fake news detection methods to real-world systems.

[0004] Compared to common domain transfer scenarios, temporal drift scenarios present greater challenges. Existing work on domain transfer in fake news detection focuses only on transfers between predefined news categories (e.g., politics, economics, health), where the boundaries between the source and target domains are relatively clear. However, datasets from different time periods often have complex temporal dependencies and blurred distribution boundaries, making the temporal drift problem more challenging. Furthermore, many existing works assume that target domain data can be obtained during the training phase, but in temporal drift scenarios, target data is generated after training is complete, making this assumption impossible. The goal of this invention is to train a model using currently available offline data, enabling the model to detect fake news on future online data (i.e., temporal generalization). A few methods improve the model's generalization on online data by learning invariant features through adversarial learning and removing domain-specific causal effects, but they do not consider the temporal pattern characteristics of news events.

[0005] From a temporal perspective, the distribution of news under the same theme exhibits certain patterns. For example, news related to the college entrance examination often concentrates during the summer. Modeling these patterns helps infer the likelihood of different news items appearing within a target time period, which can then be used to evaluate the importance of offline news data under that theme during training, thereby improving the performance of temporal generalization detection. Summary of the Invention

[0006] This invention addresses the performance degradation issue of existing fake news detection models during practical deployment due to the distribution differences between offline and online data. It proposes a fake news detection method based on time-series trend prediction, specifically including:

[0007] Extract news representations from news samples; cluster these news representations to obtain at least one topic cluster; model the news samples of each topic cluster according to their temporal distribution into a data model for the corresponding news topic; predict the temporal distribution trend of the news topic based on this data model; reweight the news samples of each topic cluster based on this temporal distribution trend and merge them into a training set; train the fake news detector using this training set; and use the fake news detector to detect fake news in the target news.

[0008] The fake news detection method of the present invention includes the step of training the fake news detector, which comprises: for news samples x in the training set... i Prediction results It is x i The probability of predicting it as fake news. The value range of is [0,1]; if Then determine x i This is fake news. Then determine x i For real news, thr is the discrimination boundary value; loss function w i,Q It is x i The weights, y i It is x i The authenticity label, if x i This is fake news, y i =1, if x i For real news, y i =0, where N is the number of news items in the training set.

[0009] In the step of predicting the time-series distribution trend in the fake news detection method of this invention, if the number of samples for a certain topic cluster is less than the reweighting threshold θ count Assign weights w to news samples of this topic cluster i,Q=1, otherwise, the time-series distribution trend of the news samples for that topic cluster is predicted, and the time-series distribution data model of the news samples is fitted by regression. If the mean absolute percentage error of the regression fit is greater than the error threshold θ, the result is considered as follows: mape Then, the news samples of that topic cluster are assigned a weight w. i,Q =1, otherwise assign a weight w to the news samples of that topic cluster. i,Q =Bound(w' i,Q );in, Bound is a reweighted computation function, representing the weighting of weights when w' i,Q >θ upper At that time, w i,Q =θ upper w' i,Q <θ lowrt At that time, w i,Q =θ lower θ lower ≤w' i,Q ≤θ upper At that time, w i,Q =w' i,Q Q' represents the news topic for which reweighting is performed; p i (f i,Q The trend of this time series distribution is represented by ).

[0010] The fake news detection method of the present invention, wherein p i (f i,Q ) = g i (f i,Q )+s i (f i,Q );g i (f i,Q ) represents the overall trend, g i (f i,Q )=k i f i,Q +m i k i k represents the growth rate. i =k+α(Q) T δ, f i,Q Where m is the frequency of news topic i in the next time period, T is the number of topic clusters used to construct the data sequence, and m is the frequency of news topic i in the next time period. i Indicates the offset, m i =m+α(Q) T γ, k, and m are initial parameters, α(Q) is the point of change for recording the growth rate and offset, δ is the scaling term, and γ is the smoothing term; s i (f i,Q This indicates a cyclical trend.

[0011] The fake news detection method of this invention is characterized by constructing a regression model of multiple news topics corresponding to a period, and obtaining s by summing the regression models. i (f i,Q ).

[0012] The fake news detection method described in this invention uses a one-pass incremental clustering algorithm to cluster the news representation. For any news sample's news representation, if the metric distance between the news representation and any existing topic cluster is greater than θ... sim If the news representation is clustered into a new topic cluster, then the news representation is clustered into a new topic cluster; otherwise, the news representation is clustered into the topic cluster with the smallest metric distance.

[0013] The fake news detection method of the present invention uses a pre-trained representation model to map the content of the news sample to a vector space to obtain the news representation.

[0014] This invention also proposes a fake news detection system based on time-series trend prediction, comprising: a representation extraction module for extracting news representations from news samples; a topic discovery module for clustering the news representations to obtain at least one topic cluster; a trend prediction module for modeling the news samples of each topic cluster into a data model corresponding to the news topic according to the time-series distribution, and predicting the time-series distribution trend of the news topic based on the data model; a reweighting module for reweighting the news of each topic cluster based on the time-series distribution trend and merging them into a training set; a detector training module for training the fake news detector with the training set; and a detection module for detecting fake news in target news using the fake news detector.

[0015] The present invention also proposes a computer-readable storage medium storing computer-executable instructions, characterized in that, when the computer-executable instructions are executed, fake news detection based on time-series trend prediction is realized as described above.

[0016] The present invention also proposes a data processing apparatus, including a computer-readable storage medium as described above, wherein when the processor of the data processing apparatus retrieves and executes computer-executable instructions in the computer-readable storage medium, it realizes fake news detection based on time-series trend prediction. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the fake news detection method based on time-series trend prediction of the present invention.

[0018] Figure 2 This is a schematic diagram of the data processing device of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0020] This invention proposes a fake news detection method based on temporal trend prediction. This method models the temporal distribution patterns of news on different topics, predicts the temporal distribution of news on different topics in the next time period, and uses the prediction results to guide the fake news detector to quickly adapt to future data distribution, thereby enabling the model to achieve better temporal generalization performance. Specifically, firstly, the training data is mapped to a vector space, and clustering is performed to discover topics; then, the temporal distribution is modeled, and a decomposable time series model is used to predict the frequency of occurrence of news corresponding to different topics in the next time period; finally, the importance of each news item in the training set to the next time period is evaluated based on the prediction results, and its impact on the model is adjusted by controlling the training loss weight of each news item during the training process, such as... Figure 1 As shown.

[0021] The fake news detection method of this invention first maps news content to a vector space to obtain its representation. Based on this representation, a clustering algorithm is used to group news samples into clusters corresponding to their topics, thereby achieving topic discovery based on news representation. Second, time-series distribution modeling and prediction are employed. The clustering results are used to model the time-series distribution of different news topics and predict the distribution of each news topic in the next time period. In the modeling process, a decomposable time series model is used to model the time-series distribution of different news topics from both the perspectives of overall trend and periodic trend. Finally, by reweighting the samples based on the time-series distribution prediction, the time-series distribution of each news topic in the next time period is predicted. The training set is reweighted at the sample level. Through reweighting, the distribution of the training set data is made closer to the next time period, thereby allowing the model trained on this dataset to better adapt to the data in the next time period and achieve better performance in the next time period.

[0022] The fake news detection method based on time-series trend prediction of the present invention includes:

[0023] Step S1: Map the content of the news samples in the training set to the vector space to obtain their representation.

[0024] Step S2: Using a clustering algorithm, news samples are grouped into clusters corresponding to their topics based on their representations.

[0025] Step S3: Model the temporal distribution of different news topics and predict their distribution in the next time period.

[0026] Step S4: Based on the distribution prediction results, the training data is reweighted at the sample level to obtain the reweighted training set.

[0027] Step S5: Train the fake news detector based on the reweighted training set.

[0028] Step S6: Use the trained fake news detector to detect fake news.

[0029] The following describes the fake news detection method based on time-series trend prediction of the present invention in further detail with reference to specific embodiments.

[0030] To better suit practical applications, this invention divides the year into four quarters, Q1-Q4, using the quarter as the unit.

[0031] Step 1: Obtain news representation

[0032] Using any existing sentence representation model (e.g., the pre-trained representation model Sentence-BERT), the news content is mapped to a vector space to obtain its representation.

[0033] Step 2: Discovering News Topics

[0034] Based on the news representation obtained in step 1, a clustering algorithm is used to assign news samples to clusters corresponding to their topics. Due to the lack of prior knowledge about the number of topics, this invention uses a single-pass incremental clustering algorithm. This algorithm does not require pre-setting the number of clusters, but rather determines the number based on a pre-set similarity threshold θ. sim Determine when to add a new cluster: For a news sample waiting to be clustered, if its similarity metric distance (the metric method is not limited here, common ones include cosine similarity) with any existing cluster is greater than θ. sim If the distance is less than a certain value, it is considered an item in a new topic and placed into a new cluster; otherwise, it is assigned to the cluster with the smallest distance.

[0035] Step 3: Temporal distribution modeling and prediction

[0036] Based on the clustering results obtained in step 2, the temporal distribution of different news topics is modeled separately, and their distribution in the next time period is predicted. To prevent topics without obvious temporal patterns from interfering with subsequent calculations, the number of news items within a cluster is less than a preset threshold θ. count News samples within a topic cluster will not participate in time-series distribution trend modeling and reweighting calculation, but will be directly assigned default weights. In this embodiment of the invention, the default weight is 1; and the number of news items within a cluster is greater than or equal to a preset threshold θ. count The news samples from the topic clusters are then used for further modeling and prediction operations.

[0037] 1. Modeling process:

[0038] Assuming there are T news topics participating in the modeling process, this invention first calculates the number of news items for each topic in different quarters. Then, it normalizes the number of news items in each quarter across topics to obtain the quarterly frequency sequence f for each topic. To model the time series distribution, this invention uses a decomposable time series model for the quarterly sequence and considers the following two trends (using topic i as an example below):

[0039] 1.1 Overall Trend

[0040] In terms of non-periodic overall trends, the frequency of a topic across different quarters (i.e., the proportion of its corresponding news items to the total number of news items in that quarter) may increase, decrease, or fluctuate slightly. To fit this variation, this invention uses a piecewise linear function:

[0041] g i (f i,Q )=k i f i,Q +m i

[0042] Where, k i =k+α(Q) T δ represents the growth rate, f i,Q It is the frequency of occurrence of theme i in quarter Q, m i =m+α(Q) T γ represents the offset. k and m are the initial parameters. α(Q) records the points of change of the growth rate and offset, δ is the scaling term, and γ is the smoothing term.

[0043] 1.2 Cyclical Trends

[0044] For themes exhibiting cyclical trends on a quarterly basis, this invention adds four additional binary regressors, corresponding to Q1 through Q4, to allow the regression models to understand the quarter to which the data points in the input sequence belong. For theme i in quarter Q, this invention obtains the quarterly seasonality function s by summing the four regression models. i (f i,Q ).

[0045] 2. Prediction process

[0046] This invention employs a regression algorithm to fit the model to the time-series distribution of different news topics from Q1 to Q-1. To predict the trend of topic i in the next quarter Q, this process superimposes the calculation results of two trend modeling functions:

[0047] p i (f i,Q ) = g i (fi,Q )+s i (f i,Q )

[0048] Step 4: Sample-level reweighting based on time distribution prediction

[0049] Based on the prediction of the topic frequency distribution in quarter Q, this invention applies sample-level reweighting to the training set, so that the model trained on the reweighted training set can better adapt to the data in the next time period, quarter Q.

[0050] First, for regression fitting processes where the mean absolute percentage error (MAPE) is greater than a preset threshold θ mape News samples on the topic are not reweighted; instead, a default weight of 1 is assigned. For news items less than or equal to θ... mape Only news samples corresponding to the news topic i are reweighted. For all news samples that undergo reweighting, this invention calculates and normalizes the ratio p between the predicted frequencies of topic i. i (f i,Q ) and the sum of all predicted frequencies for the retained topic:

[0051]

[0052] Here, Bound is a function used to constrain the range of the calculated weights. To avoid instability during training, this invention will reweight the corresponding w' in the representation. i,Q The value is less than the preset threshold θ lower and greater than the preset threshold θ upper The weights are set to θ respectively. lower or θ upper That is, for news samples that undergo reweighting, when w' i,Q >θ upper At that time, w i,Q =θ upper When w' i,Q <θ lower At that time, w i,Q =θ lower When θ lower ≤w' i,Q ≤θ upper At that time, w i,Q =w' i,Q Q' represents the news topic for which reweighting is performed.

[0053] All news samples are merged to obtain the training set required for training the fake news detector of this invention. The training set consists of three parts: 1. News items within a cluster with a number of items less than a preset threshold θ. count News samples within a topic cluster, with weight w i,Q=1; 2. During the regression fitting process, MAPE is greater than the preset threshold θ. mape The news sample of the news topic, its weight w i,Q =1; 3. The news samples undergoing reweighting calculation have weights w i,Q =Bound(w' i,Q ).

[0054] New weights w for samples of topic i in the training set i,Q This corresponds to the prediction of the frequency of news about the topic i appearing in the future over a period of time Q. If the predicted frequency of topic i is lower than its frequency in the training set, the value will be less than 1, and the sample of that topic will be weighted less; conversely, if the predicted frequency is higher than its frequency in the training set, the value will be greater than 1, and the sample will be weighted more.

[0055] Step 5: Train the fake news detector

[0056] The method of this invention is compatible with any neural network-based fake news detector. Here, the BERT model is used as an example. Specifically, given a sample x... i This invention combines the special word character [CLS] and x i Connect them together and input them into BERT. The output of [CLS] represents o i It is then fed into a multilayer perceptron (MLP) with a sigmoid function for final prediction:

[0057]

[0058] in, This represents the probability of it being predicted as fake news, with a value ranging from [0,1]. If Then determine x i It is fake news; otherwise, it is judged as x. i For real news, thr is a preset discrimination boundary value.

[0059] The difference between this invention's method and conventional methods lies in the use of new weights derived from time distribution predictions to increase or decrease the impact of individual news samples during backpropagation. This invention minimizes the weighted cross-entropy loss function L during training:

[0060]

[0061] Where w i,Q It is sample x i The new weight, y i It is its authenticity label (0 for real news, 1 for fake news), and N is the size of the training set.

[0062] Figure 2This is a schematic diagram of the data processing apparatus of the present invention. Figure 2 As shown, embodiments of the present invention also provide a computer-readable storage medium and a data processing apparatus. The computer-readable storage medium of the present invention stores computer-executable instructions. When these computer-executable instructions are executed by the processor of the data processing apparatus, they achieve fake news detection based on time-series trend prediction. Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware (e.g., processor, FPGA, ASIC, etc.), and the program can be stored in a readable storage medium, such as a read-only memory, a disk, or an optical disk. All or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module in the above embodiments can be implemented in hardware, for example, by implementing its corresponding function through an integrated circuit, or it can be implemented in the form of a software functional module, for example, by a processor executing a program / instruction stored in memory to implement its corresponding function. Embodiments of the present invention are not limited to any particular combination of hardware and software.

[0063] The fake news detection method based on time-series trend prediction of the present invention can obtain clustering results that group news into different topics by discovering topics based on news representation; model the time-series distribution of different news topics by clustering the clustering results and predict the distribution of different topics in the next time period; in addition, sample-level reweighting is applied to the training set to make the distribution of the training set closer to the next time period, thereby enabling the model trained on it to achieve better detection results in the next time period.

[0064] The above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the present invention, and the patent protection scope of the present invention should be defined by the claims.

Claims

1. A method for detecting fake news based on time-series trend prediction, characterized in that, include: Extracting news representations from news samples; Cluster the news representation to obtain at least one topic cluster; By distributing news samples from each topic cluster in a time series, a data model is created for the corresponding news topic. Based on this data model, the time series distribution trend of the news topic is predicted. Based on this time-series distribution trend, news samples from each topic cluster are reweighted and then merged into a training set; where, if the number of samples in a certain topic cluster is less than the reweighting threshold... Assign weights to news samples of this topic cluster. Conversely, if the trend is not specified, a time-series distribution trend prediction is performed on the news samples of that topic cluster, and a regression model is used to fit the time-series distribution data of the news samples. If the mean absolute percentage error of the regression fit is greater than the error threshold, the prediction is considered complete. Then, the news samples of that topic cluster are assigned weights. Conversely, news samples from that topic cluster are assigned weights. ; , For reweighted computation functions, it means hour, , hour, , hour, ; News topics for reweighting calculation; This represents the trend of the time series distribution; The fake news detector was trained using this training set; This fake news detector is used to detect fake news in target news.

2. The method for detecting fake news as described in claim 1, characterized in that, The steps for training this fake news detector include: For the news samples in this training set Prediction results , yes The probability of predicting it as fake news. The value range of is [0,1]; if Then determine This is fake news. Then determine This is real news. It is to determine boundary values; loss function ; yes The weight, yes The authenticity label, if This is fake news. ,like This is real news. N is the number of news items in the training set.

3. The method for detecting fake news as described in claim 2, characterized in that, ; As the overall trend, , Indicates the growth rate. , It is a news topic The frequency of occurrence in the next period The number of topic clusters to construct the data sequence, Indicates the offset. , and These are the initial parameters. To record the points of change in growth rate and offset, δ is the scaling adjustment term and γ is the smoothing term; It is a cyclical trend.

4. The method for detecting fake news as described in claim 3, characterized in that, Construct a regression model for multiple news topics corresponding to the cycle, and obtain the result by summing the regression models. .

5. The method for detecting fake news as described in claim 1, characterized in that, The news representation is clustered using a one-pass incremental clustering algorithm. For any news representation of a news sample, if the metric distance between the news representation and any existing topic cluster is greater than a certain value, then the clustering is considered successful. If the news representation is clustered into a new topic cluster, then the news representation is clustered into a new topic cluster; otherwise, the news representation is clustered into the topic cluster with the smallest metric distance.

6. The method for detecting fake news as described in claim 1, characterized in that, Using a pre-trained representation model, the content of the news sample is mapped to a vector space to obtain the news representation.

7. A fake news detection system based on time-series trend prediction, characterized in that, include: The characterization extraction module is used to extract news characterizations from news samples. The topic discovery module is used to cluster the news representation to obtain at least one topic cluster; The trend prediction module is used to model the news samples of each topic cluster into a data model of the corresponding news topic according to the time-series distribution, and based on the data model, predict the time-series distribution trend of the news topic. The reweighting module is used to reweight the news from each topic cluster based on the time-series distribution trend and then merge them into a training set; where, if the number of samples in a certain topic cluster is less than the reweighting threshold... Assign weights to news samples of this topic cluster. Conversely, if the trend is not specified, a time-series distribution trend prediction is performed on the news samples of that topic cluster, and a regression model is used to fit the time-series distribution data of the news samples. If the mean absolute percentage error of the regression fit is greater than the error threshold, the prediction is considered complete. Then, the news samples of that topic cluster are assigned weights. Conversely, news samples from that topic cluster are assigned weights. ; , For reweighted computation functions, it means hour, , hour, , hour, ; News topics for reweighting calculation; This represents the trend of the time series distribution; The detector training module is used to train the fake news detector using this training set; The detection module is used to detect fake news in the target news using the fake news detector.

8. A computer-readable storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, the fake news detection based on time-series trend prediction as described in any one of claims 1 to 6 is implemented.

9. A data processing apparatus comprising the computer-readable storage medium of claim 8, wherein when a processor of the data processing apparatus retrieves and executes computer-executable instructions in the computer-readable storage medium, it implements fake news detection based on time-series trend prediction.

Images