Reading of false news detection papers (3): semi supervised content-based detection of misinformation via tensor embeddings

Paper title ：Semi-supervised Content-based Detection of Misinformation via Tensor Embeddings
Source of the paper ：IEEE 2018

# Based on news text 、# Semi supervision 、# The graph structure

One 、 The motivation of the article

Most of the existing work focuses on manual feature extraction and supervised classification models using a large number of labeled articles . But in reality, there are often not many labeled data , There are only a few labels .

Two 、 Main work

The author simulates the above situation , Express the set of news articles as multidimensional tensors , Use tensor decomposition to derive concise article embedding , The space used to capture each article / Context information , And use these embeddings to create article by article graphs , Spread Limited labels on the graph .
Using tensor representation and decomposition of news articles , Carefully constructed KNN Graph and propagate limited tagged article information to reason about a larger set ：
（1） The graph representation of news articles is derived by tensor based article embedding ;
（2） Develop semi supervisory methods , Propagate known tags on the graph to identify unknown tags ;
（3） Collect a large number of error information and real news article data sets publicly shared on social media ;
（4） Evaluate the proposed method on real data sets .

1. Express the content of the article as tensor （Tensor decompositon）
Two different tensor construction methods are proposed ：
（1） Frequency based tensor ： A the third mock examination tensor is constructed (word, word, news), For each news article, a non-zero co-occurrence matrix is created
（2） Tensor based on binary ： Construct a tensor (word, word, news）, The entries used and appeared are Boolean values , Instruction No i And the first j Whether words appear at least once in the predefined window .

2. News articles KNN chart
Use steps 1 The tensor embedding calculated in constructs the graphical representation of news articles , Use factor matrix C Building news articles KNN chart G,C Each column in is the representation of the corresponding news article in the potential theme space , therefore , By means of C Build a KNN chart , We can find similar articles in this space .
Specific operation ：
（1） take C Each line in is treated as R A point in dimensional space ;
（2） Calculate the... Between news $l_2$ Distance and find C At every point in k- More recently 【 The number of news articles is very large in practice , We can take advantage of kd-tree（k Dimension tree ） In order to find more efficiently for each article k Nearest neighbor 】;
（3）G Each node in represents a news article , Each edge encodes two articles that are similar in the embedded space ;
Final plan G Is an undirected symmetric graph , Each of these nodes is connected to at least k Nodes . The compact representation of the figure is M×M Adjacency matrix .

3. Belief Propagation
Using a belief propagation algorithm that assumes homogeneity .
Use fast and linear FaBP variant ：
（1） The algorithm is proved to be insensitive to the size of known tags , Therefore, we believe that FaBP Good learning performance can be achieved by using only a few known tags .
（2）FaBP stay G The number of edges in is linear , Therefore, it can be used for large-scale graphics .

3、 ... and 、 Data sets

1、Political（150 A political news article ）;
2、Bulgarian（68 Real news and 69 Fake news ）.