Paper with Mariana Teles on using AI text analysis to classify and summarize open-ended student feedback in large classes. We describe a practical guide for instructors who want to apply sentiment analysis and topic modeling to course evaluations without writing code. Published in College Teaching.
We published our transforEmotion paper in Computational Communication Research. The R package solves modal fragmentation by unifying text, image, and video emotion recognition in a single tool. It supports zero-shot classification with custom emotion labels and runs locally without API dependencies.
Presented "transforEmotion: Multimodal Emotion Analysis in R" at the IAAP Webinar: AI & Emotion in Digital Life. I walked through text, image, and video emotion workflows in the package.
Presented "transforEmotion in Practice: Multimodal Emotion Analysis in R" at the Vanderbilt University QMDA Colloquium. I demonstrated transforEmotion workflows on real data, covering text sentiment, video frame-by-frame emotion extraction, and dynamic network visualization of emotion time series.
Preprint on Reddit deplatforming and toxicity dynamics on arXiv. We analyzed 6 communities (funny, gaming, gifs, pics, videos, technology) from the MADOC dataset across 4 Reddit ban waves (2015-2020). We found two migration regimes: a hostile takeover phase and a toxic equilibrium where fewer than 5% of newcomers reached central positions, yet toxicity doubled. Code on GitHub.
Preprint on validating LLM agent social simulations on arXiv. We used YSocial with Dolphin 3.0 (uncensored Llama 3.1 8B) to simulate Voat's v/technology over 30 days and compared network structure, toxicity distributions, topic modeling (BERTopic), and semantic similarity against real data. Code and 14+ analytical scripts on GitHub.
We tested measurement invariance of the Youth Externalizing Problems Screener (YEPS) across 32 countries using multi-group confirmatory factor analysis. Published in Development and Psychopathology.
A large cross-national study examining gender and country-level differences in adolescent emotional experiences across 32 countries. We used multi-group measurement invariance testing to compare positive and negative affect. Published in Applied Research in Quality of Life.
We used European Social Survey Round 10 data from 31 nations and identified three dimensions of democratic attitudes: liberal, populist, and social-democratic. Partial measurement invariance held across countries. Workers showed stronger support for populist and social-democratic ideals than service professionals. Published in Sociologija.
We introduced a network-based approach to test measurement invariance of the Short Dark Triad, moving beyond traditional CFA methods. We compared 14 cultures. Published in Personality and Individual Differences.
We used two-wave longitudinal data to test whether AIQ-IV identity dimensions predict changes in well-being and internalizing/externalizing problems over time. Published in Journal of Child and Family Studies.
We tested configural, metric, and scalar invariance of the Aspects of Identity Questionnaire across 30 countries. This is the largest cross-cultural identity structure study to date. Published in Journal of Research on Adolescence.
We tested whether collective ("we") identity orientations predict adolescent well-being beyond what basic psychological needs explain. We used structural equation modeling to separate the unique contribution of collective identity. Published in Identity.
We analyzed 220 YouTube videos of leaders from 15 countries using deep learning (frame-by-frame detection of 6 emotions + neutral). We found significant differences in negative emotion expression between populist and non-populist leaders. Code on GitHub. Published in Journal of Computational Social Science.
Network analysis mapping associations between Machiavellianism, narcissism, psychopathy, and domains of emotional functioning. We used EBICglasso network estimation. Data and code on OSF. Published in International Journal of Psychology.
We released the Multi-Platform Aggregated Dataset of Online Communities (MADOC) on Zenodo, with Python (pymadoc) and R (rMADOC) packages. MADOC covers Reddit, Voat, Bluesky, and Koo platforms with communities including CringeAnarchy, fatpeoplehate, funny, gaming, and gifs in Parquet format.
Published "Statistički metodi: Praktični pristup u R-u" (Statistical Methods: A Practical Approach in R). The textbook covers core statistical methods with R code examples and practical exercises. In Serbian.
Released transforEmotion v1.5 on GitHub and CRAN. This version added Python environment provisioning via uv, improved video frame extraction, and streamlined reticulate setup for calling transformer models from R.
We applied Dynamic Exploratory Graph Analysis to 220 political leader videos and found a two-dimensional network separating positive from negative emotions. Populist leaders showed less connected anger expressions. We used transforEmotion for zero-shot video classification. Code on GitHub. Published in advances.in/psychology.
Updated preprint on facial emotional expressions of political leaders on arXiv. This version adds a model comparison study evaluating CLIP variants against CNN-based models for political emotion annotation. Annotation data and code on GitHub.
Presented on core-periphery structure and trust dynamics in online communities at Sunbelt 2024. I compared Reddit and Voat time series data. Slides and code on GitHub.
We examined structural validity and gender invariance of the AIQ-IV among Serbian adolescents, and tested relationships with mental health and basic psychological needs. Published in Journal of Personality Assessment.
Taught a Python workshop at the CoPA Training School. The course covered pandas, data wrangling, visualization, and basic NLP for social science data analysis. Materials online.