SOCIAL MEDIA BOT ACCOUNT DETECTION MODEL BASED ON MULTI-VIEW METHOD WITH ATTENTION MECHANISM

Abstract

The article proposes a multi-view model for detecting bot accounts, which is based on the use of the ensemble Gradient Boosting method and the attention mechanism for integrating the results of the analysis of individual representations. The proposed approach involves considering each type of data as a separate representation, for which its own feature vector is formed, and a local model is trained. The behavioral representation considers the temporal characteristics of user activity, the attributive representation - the properties of the profile, the content representation - the textual features of posts using semantic models, and the visual representation - the features of images, including metadata analysis and OCR. For each representation, a local estimate of the probability of automated activity and a data quality indicator are calculated. The integration of results is performed using an attention mechanism that determines the weight of each representation depending on its informativeness and reliability. This allows us to adaptively take into account incompleteness or heterogeneity of data and increases the model's resilience to the absence of individual types of information. The final score is formed based on a weighted combination of representations using the Gradient Boosting metamodel. A feature of the proposed approach is the ability to interpret the results by assessing the contribution of each representation and determining the level of confidence of the model. This ensures transparency of the decisions made and allows the model to be used in practical social network analysis systems. The proposed approach extends existing bot detection methods by combining multi-species learning, adaptive aggregation, and data quality consideration, which increases the efficiency of the developed system.