Toward Uncertainty-Aware Influence Detection on Twitter using LDA and Dempster–Shafer Theory
This study proposes a novel framework for identifying the dominant domain of influence of social media actors on Twitter under conditions of uncertainty. Existing influence analysis approaches rely on network topology, popularity indicators, or content analysis separately and often fail to explicitly manage uncertainty and conflicting evidence. To address this research gap, the proposed framework combines Latent Dirichlet Allocation (LDA), Natural Language Processing (NLP), Belief Function Theory and Dempster’s rule of combination within a unified evidential reasoning framework. First, NLP preprocessing techniques including tokenization, stop-word removal, lemmatization and text normalization are applied to transform raw tweets into a structured textual corpus. LDA is then employed to extract latent thematic topics and estimate topic distributions. Subsequently, thematic relevance and engagement indicators are transformed into belief mass functions. These heterogeneous sources of evidence are fused through Dempster–Shafer theory and pignistic probabilities are computed to identify the dominant influence domain. Experimental validation conducted on two influential public figures from distinct domains demonstrates that the proposed framework successfully identifies dominant influence domains while explicitly accounting for uncertain evidence. The results show strong consistency between thematic relevance and audience engagement indicators, confirming the robustness of the proposed evidential fusion strategy. The main novelty of this work lies in transforming topic-modeling outputs into belief mass functions and integrating them with engagement-based evidence through Dempster’s combination rule. The proposed framework provides a rigorous, interpretable and uncertainty-aware solution for social media influence-domain identification and offers practical applications in influencer marketing, digital communication and reputation management.