Method for detecting sources of disinformation based on ensemble machine learning models.

The presented study developed a method for detecting sources of disinformation based on ensemble machine learning models. Modern methods of combating disinformation and detecting false content were analyzed. A fake news identification system based on the ensemble approach was implemented as part of the work, and its architectural structure was described. The main stages of cleaning text data obtained from social networks and news are described in detail, in particular, the normalization of categorical variables. Statistical analysis of the text and analysis of the criteria for identifying sources of disinformation distribution are carried out. An analysis of the balance of target and auxiliary variables was performed, which made it possible to identify dependencies between the language of the message and reliability. Two types of text embeddings and corresponding classification models were used for modeling: linear regression and logistic regression. The final stage was the application of an ensemble of models, which made it possible to combine the predictive ability of both models. The results showed that the combination of approaches improves classification quality, especially in conditions of unbalanced data. Using an ensemble of models allowed us to increase the accuracy from 73% (model 1) and 71% (model 2) to 78%.