Real Talk, Virtual Faces: A Formal Concept Analysis of Personality and Sentiment in Influencer Audiences
Shahram Chaudhry, Sidahmed Benabderrahmane, Talal Rahwan · 2026
Read Real Talk, Virtual Faces: A Formal Concept Analysis of Personality and Sentiment in Influencer Audiences by Shahram Chaudhry, Sidahmed Benabderrahmane, Talal Rahwan (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-24410.txt.
Introduction
Chaudhry and co-authors compare how audiences talk about virtual and human influencers on YouTube. Surveys and sentiment counts show what people say, but not how affect, topics, and language cues are organised together. Their framework pairs formal concept analysis and rule mining with semantic embeddings of the concepts found. Across 69,498 comments from three matched pairs, human influencer talk centres on one stable pattern. Virtual influencer talk splits into several regimes, spreads wider, and holds more negative sentiment on sensitive topics.
Catalog record
| CREATOR | Shahram Chaudhry, Sidahmed Benabderrahmane, Talal Rahwan |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 33076 |
| LICENSE | http://creativecommons.org/licenses/by-sa/4.0/ |
| RIGHTS | CC BY SA |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.24410
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Real Talk, Virtual Faces: A Formal Concept Analysis of Personality and Sentiment in Influencer Audiences Shahram Chaudhry, Sidahmed Benabderrahmane, Talal Rahwan Abstract Virtual influencers (VIs)—digitally synthetic social-media personas—attract audiences whose discourse appears qualitatively different from discourse around human influencers (HIs). Existing work characterises this difference through surveys or aggregate engagement statistics, which reveal what audiences say but not how multiple signals co-occur. We propose a two-layer, structure-first framework grounded in Formal Concept Analysis (FCA) and association rule mining. The first layer applies FCA with support-based iceberg filtering to weekly-aggregated comment data, extracting discourse profiles—weekly co-occurrence bundles of sentiment, Big Five personality cues, and topic tags. The second layer mines association rules at the comment level, revealing personality–sentiment–topic dependencies invisible to frequency-table analysis. Applied to YouTube comments from three VI–HI influencer pairs, the two-layer analysis reveals a consistent structural divergence: HI discourse concentrates into a single, emotionally regulated (stability-centred) regime (low neuroticism anchoring positivity), while VI discourse supports three structurally distinct discourse modes, including an appearance-discourse cluster absent from HI despite near-equal marginal prevalence. Topic-specific analyses further show that VI contexts exhibit negative sentiment in psychologically sensitive domains (mental health, body image, artificial identity) relative to HI contexts. Our results position FCA as a principled tool for multi-signal discourse analysis and demonstrate that virtuality reshapes not just what audiences say, but the underlying grammar of how signals co-occur in their reactions. 1 Introduction Formal Concept Analysis (FCA) provides a mathematically principled framework for discovering the co-occurrence structure hidden in binary object–attribute relations [ 4 ] . FCA’s lattice of formal concepts identifies maximal closed sets of attributes shared by non-trivial subsets of objects—a form of structure that cannot be read off from marginal frequencies or independent feature distributions. This property makes FCA particularly powerful when the scientific question concerns how signals co-occur rather than how frequently individual signals appear. Why FCA here. Our contribution is not merely applying FCA to a new domain: we show that matched marginal prevalences can yield qualitatively different closed-set and rule grammars , and we package this as a reusable structural diagnostic (marginals \rightarrow iceberg concepts \rightarrow rule-cluster comparison) for multi-signal discourse data. While FCA has been widely applied in knowledge discovery and text analysis, its potential as a structural diagnostic for large-scale social discourse data remains underexplored. We ask: does the digitally synthetic nature of virtual influencers leave a detectable imprint in the lattice structure of their audiences’ discourse, beyond what marginal statistics reveal? Virtual influencers (VIs) such as Lil Miquela, APOKI, and Milla Sofia have amassed millions of followers and major brand partnerships, yet audience responses to them remain emotionally complex and at times unstable. While some users express curiosity or admiration, others respond with discomfort or skepticism—especially on sensitive topics such as mental health or body image [ 3 , 7 ] . Capturing these reactions and their interactions with sentiment and personality signals requires tools that model structure , not only distributions. Although topic_appearance has near-equal marginal prevalence in VI and HI, it participates in multiple VI rule clusters and in none of the HI rule sets, illustrating why structural analysis is needed beyond frequency tables. Contributions. We introduce a two-layer structure-first framework combining (i) weekly-aggregated FCA with support-based iceberg filtering, which extracts recurring audience discourse profiles across time, and (ii) comment-level association rule mining, which yields directional personality–sentiment–topic regularities. Concretely, we: 1. Show that VI and HI discourse produce qualitatively different weekly FCA concept structures: HI-only concepts are sentiment-centred (positive sentiment, topic positivity, low neuroticism), while VI-only concepts are personality-centred (high openness, high agreeableness, low conscientiousness). 2. Demonstrate that under identical rule-mining thresholds, VI discourse yields 51 filtered rules versus 8 for HI, and that VI rules form three structural clusters including an appearance-discourse mode absent from HI. 3. Show that marginal prevalence equality can coexist with lattice-level structural asymmetry—the core methodological claim motivating FCA. 4. Document that VI contexts show a higher concentration of negative sentiment in psychologically sensitive topics (artificial identity, body image, mental health) relative to HI contexts. Research questions. RQ1. (Stable profiles.) How do weekly FCA discourse profiles differ between VI and HI conditions? RQ2. (Rule grammar.) What directional association rules characterise VI vs. HI comment-level discourse, and how do the rule sets differ in size and structure? RQ3. (Topic sensitivity.) How do sentiment and personality patterns differ between VI and HI audiences across psychologically sensitive topics? FCA perspective. Beyond the empirical analysis of influencer discourse, this work illustrates how FCA can serve as a structural diagnostic for multi-signal social data: marginal attribute distributions may appear similar across conditions, yet their closed-set organisation and rule bases can differ substantially. The comparison of concept lattices and derived rule clusters therefore provides a principled method for detecting latent structural differences in discourse data. In this sense, the contribution is methodological as well as empirical: we demonstrate that FCA-based closed-set structure can serve as a diagnostic signal when marginal attribute frequencies alone fail to distinguish between discourse regimes. 2 Related Work 2.1 Formal Concept Analysis: theory and applications FCA was introduced by Ganter and Wille [ 4 ] as a lattice-theoretic framework for conceptual data analysis. Kuznetsov and Obiedkov [ 6 ] analyse computational complexity and early ML applications; Stumme et al. [ 13 ] link FCA to frequent closed itemsets, showing that the closed-set family is an order-of-magnitude smaller than all frequent itemsets yet losslessly condensed. Pasquier et al. [ 10 ] demonstrate efficient closed-set mining, and Poelmans et al. [ 11 ] survey FCA applications across knowledge discovery, text mining, and social analytics. Our work extends this line by constructing multi-signal formal contexts from psycholinguistically annotated comment corpora at two granularities (weekly and comment level). 2.2 Association rule mining Agrawal and Srikant [ 1 ] introduced Apriori for frequent itemset and association rule mining. The Duquenne–Guigues basis [ 4 ] provides an irredundant implication set derived from FCA’s formal context. Our use of the Galicia platform [ 15 , 16 ] grounds rule generation in FCA’s approximate closed sets rather than heuristic frequency enumeration. 2.3 Virtual influencer audiences Research on VI audiences consistently emphasises authenticity, parasocial ties, and realism [ 14 , 5 ] . Batista and Chimenti [ 3 ] show that highly humanised VIs evoke emotionally complex reactions from admiration to skepticism. Arsenyan and Mirowska [ 2 ] demonstrate that disclosure of virtual status moderates trust and purchase intent; Xie-Carson et al. [ 17 ] document uncanny-valley effects in tourism contexts; Sands et al. [ 12 ] and Lou et al. [ 8 ] extend this to parasocial bonds and appearance evaluations. Looi et al. [ 7 ] analyse 48,000 Instagram comments on Lil Miquela, finding polarised emotional reactions driven by identity ambiguity. Yan et al. [ 18 ] show that hyper-realistic VI designs weaken emotional attachment relative to stylised designs. Existing studies rely on surveys or aggregate metrics; we complement them by analysing actual comment text at scale with structure-discovery tools. 2.4 Psycholinguistic features in social media Big Five personality proxies inferred from text have been applied to polarisation and community detection [ 9 ] . We move beyond independent application by building formal contexts that encode sentiment, personality, and topics jointly, so FCA can