| CPC G06Q 50/01 (2013.01) [G06Q 30/0201 (2013.01)] | 18 Claims |
|
1. A method comprising:
transmitting a query over a network to a search Application Programming Interface (search API) provided by a social media platform;
determining, in response to the query, a count of a quantity of articles posted by each of a plurality of social media users toward articles based on information lacking a high degree of evidential basis and articles based on information having a high degree of evidential basis;
classifying each of the plurality of social media users into one of a plurality of social media user communities based on their respective count of the quantity of articles posted by each social media user toward information lacking a high degree of evidential basis and information having a high degree of evidential basis;
identifying user profile features representative of users in each of the plurality of social media user communities;
configuring an auto-encoder of a fake news detection model using sequence-to-sequence learning to encode the articles by providing as input to the auto-encoder, each word of a respective article and a latest hidden state of a Long Short-Term Memory (LSTM) neural network providing a latent representation of the LSTM neural network;
configuring a decoder of the fake news detection model to generate an attempted reconstruction of the respective article using as input to the decoder, the latent representation of the LSTM neural network;
detecting, by the decoder of the fake news detection model, the respective article posted on social media is either (i) one with information having a high degree of evidential basis or alternatively, (ii) one with information lacking a high degree of evidential basis based on a combination of the user profile features of the user that posted the respective article and the latent representation provided as input to the LSTM neural network;
calculating a credibility score for a subset of the social media users on the basis of user-adjacency and utilizing only the subset of the social media users having directly posted the respective news article or having reposted the respective article without adding comments;
clustering, by the fake news detection model, each of the subset of the social media users into clusters according to the user-adjacency of each of the respective ones of the subset of the social media users having user profile features having sharing similarities with others within the subset of the social media users;
weighting, by the fake news detection model, each of the clusters according to cluster size, wherein the social media users within larger clusters correspond to a higher credibility score and wherein the social media users within smaller clusters correspond to a lower credibility score; and
outputting, by the fake news detection model, a classification for the respective article as either (i) the respective article as one with information having a high degree of evidential basis or alternatively, (ii) the respective article as one with information lacking a high degree of evidential basis according to the credibility score calculated for the social media user having posted the respective article to social media.
|