Reading papers on false news detection (5): a semi supervised learning method for fake news detection in social media

Paper title ：A Semi-supervised Learning Method for Fake News Detection in Social Media
date ：IEEE2020

# Based on news text （+ Image multimode ）、# Semi supervision 、# Pseudo label 、#LDA

One 、 The basic content

utilize LDA Method to fake label the unlabeled data , So as to better train unmarked CNN Model , Good results have been achieved .

Two 、 Main work

SLD-CNN
Based on semi supervised learning framework , Use CNN For tagged and unlabeled data .
（1） use first CNN Extract various features of text and image data ;
（2） Using linear discriminant analysis （LDA） Predict the categories of unclassified data ;
（3） A method is proposed to calculate the fitness function , To improve the effect of prediction categories in each step .

3、 ... and 、 Model framework

• CNN Need to tag data to optimize the network , And unmarked data cannot be in CNN Use in . Therefore, the use is based on LDA Method to predict unlabeled data .
• LDA Linear discriminant analysis ：
  A concept close to variance analysis and regression analysis . In every statistical method , The dependent variable is modeled as a combination of other variables . But in variance analysis and regression analysis , The dependent variable is the distance type , And in the LDA in , Dependent variables are nominal or ordered .
  The paper assumes that each features Can be modeled as a A random variable .
  Covariance matrix is a general form of variance of numerical variables in different directions , And because variance represents the distribution of values of random variables near the mean . therefore ,$n$ Covariance matrix of variables Express Around the mean vector $n$ Probability distribution in dimensional space .
  If there is $n$ Random variables $\{h_1,\dots ,h_n\}$ Make each variable contain $m$ An example （ And stored in dimension $m\times n$ Matrix $D$ in , Among them the first $i$ Column number 1 $j$ The element of the row represents $x_i$ and $y_i$ The covariance between ）.
  $\sum [i,j]=Cov(x_i,y_i)$
  $Cov(x_i,y_i)=\frac{1}{m}{\sum }_{l=1}^m[(D(l,i)-{\mu }_i)(D(i,j)-{\mu }_j)]$
  ${\mu }_i$ and ${\mu }_j$ Is the second of the matrix $i$ That's ok 、 The first $j$ The variable mean of the column , according to Fisher Linear discriminant theory , When the difference between the mean values is the largest , The difference between the two parties is the smallest , The discrimination effect is the best ：
  $W=({\sum }_0+{\sum }_1{)}^{-1}({\overline{\mu }}_1)-({\overline{\mu }}_0)$
  Due to the problem 2 Categories （ Real articles and fake articles ）, So the index is 0 and 1.
  Because for every class , The mean is the vector of each mean of each feature . This means that the two class separation levels are at their maximum , We can reduce the dimension （ That is, the number of features ） And retain the important features . This process applies to data without labels .
• Semi supervision LDA（SLDA）
  SLDA It's an iterative algorithm , Get the number of features and classes and iteratively maximize the discrimination class （ Two classes ：fake—0、real—1）, therefore SLDA The output of is between 0 and 1 Number between .
  
• combination CNN
  There may be errors in estimating the value of the tag , In order to control its effect on CNN The impact of the classification process , The influence factor is introduced α：
  ${\theta }^{\ast }=argmin[L_{labeled}(\hat{y},y)+\alpha L_{unlabeled}(\hat{y},y)]$

【 The activation functions of the first and second convolution layers are ReLU, The activation function of the output layer is Softmax, Use Adam Optimizer 】
Parameter information used ：

Four 、 Data sets

Multimode （ Text 、 Images ） Data sets come from well-known websites ：20015 News articles , among 11941 Article is false ,8074 This article is correct .