A two-layer model to detecting falsified information using neural networks in socially oriented systems

Abstract

The subject matter of the article is the problem of detecting fabricated information in socially oriented systems characterized by significant user load. The goal of the work is to develop of a two-layer fake information classification model based on a combination of a naive Bayesian classifier and a hybrid recurrent-convolutional neural network. The following tasks were solved in the article: conducting expert evaluation and domain analysis to determine basic classes of fake information; analyzing linguistic markers of disinformation and developing feature vectors for classification; developing models for data segregation using a naive Bayesian classifier; conducting experimental verification of the proposed two-layer model in comparison with the RCNN approach. The following methods used are – analytical method for forming a set of disinformation markers; inductive method for determining the target set of indicators for implementing the second layer of the model; expert evaluation for determining the most influential efficiency factors and feature weight coefficients; experimental and multi-criteria evaluation methods for determining the most effective model. The following results were obtained – a classification structure for types of fake information was formed, including five categories from jokes to globally harmful news. A set of discriminative features characteristic of fabricated information was developed, including primary linguistic markers and secondary stylometric indicators. It was determined that the approach using a two-layer model demonstrated, on average, a 15% improvement in efficiency compared to direct application of a hybrid recurrent-convolutional neural network. Conclusions: the application of a two-layer data classification model successfully expands the capabilities of basic detection of data falsification, including scale assessment and analysis of fabrication intentionality. Empirical analysis shows that implementation of a two-layer model with a naive Bayesian classifier achieves an average 15% performance improvement compared to simple neural network application. This performance difference becomes particularly significant in high-throughput systems where rapid identification and response to fabricated information are critical operational parameters. The obtained result allows us to assert the feasibility of implementing the proposed approach, and accordingly, provides the opportunity to reduce the impact of such information in socially oriented systems, especially during crisis situations.