Affective computing scholarship and the rise of China: a view from 25 years of bibliometric data

The paper investigates how affective computing research has evolved over 25 years, who the major contributors are, and how international collaborations have formed, with special attention to the rise of China relative to the US and Europe. The authors motivate the study by outlining the growing ubiquity and commercial penetration of emotional AI across education, entertainment, transportation, workplace monitoring, and public governance, alongside ethical, legal, cultural, and technical controversies (bias, accuracy, privacy, and cross-cultural validity). They highlight divergent regulatory landscapes (e.g., EU’s proposed AI regulation versus weaker protections in parts of the Global South and certain Asian workplaces shaped by Confucian norms), and the need for a synoptic quantitative mapping given rapid publication growth. The research question centers on mapping production trends, key national hubs, collaboration networks, and thematic shifts in affective computing over time to understand geopolitical dynamics (US–China), interdisciplinarity, and commercialization trajectories.

The background situates affective computing as a multidisciplinary domain coined by Picard (1995) and now widely deployed as emotional AI. Prior work documents pervasive applications (e.g., ClassDojo, Moxie, Spotify, driver monitoring) and workplace adoption alongside debates over algorithmic management’s benefits and harms (productivity vs. technostress, dignity, autonomy). Studies show bias and inaccuracy in facial biometrics across gender and race, and risks in cross-cultural deployment without dataset adaptation. Regulatory literature contrasts stringent European data protection proposals with weaker protections in parts of Asia and the Global South, where cultural norms and legal gaps facilitate broader data collection. The paper also reviews the evolution and utility of bibliometrics/scientometrics, open science tools (bibliometrix, VOSviewer), and prior bibliometric mappings in related AI domains. The theoretical landscape of emotions is reviewed later in the discussion, contrasting Ekman’s basic emotions with constructionist models (Barrett), and advances in text-based affect models (OCC, Hourglass, SenticNet), highlighting implications for dataset design and modality-specific methods.

Data source: Web of Science (SCI-EXPANDED, SSCI, A&HCI, CPCI-S, CPCI-SSH, BKCI-S, BKCI-SSH, ESCI). Search term: Topic = “affective computing”. Timespan: 1995–2020 (reflecting Picard’s coinage). Language: English. Query date: April 5, 2021. Initial results: 3527; refined to English: 3469; after removing 2021 early access items and duplicates: 3448 items (journals, conference proceedings, reviews, books/chapters, editorials). Analysis tools: R bibliometrix package and Biblioshiny for source, author, and document-level analyses; conceptual (co-word), intellectual, and social (co-authorship) structure mapping; network centrality (betweenness, closeness) and thematic evolution via co-occurrence of author keywords (top 1000, min 3 occurrences). Descriptive metadata (authorship counts, collaboration indices, citations) and temporal segmentation into three periods (1995–2005; 2006–2015; 2016–2020) for comparative analyses of output, citations, and collaboration networks.

- Output growth: Annual growth rate of affective computing publications is 11.36%. Publications in 2016–2020 (n=1660) nearly equal the previous 20 years combined (n=1726). - Corpus statistics: 3448 documents; 1908 sources; 88,875 references; average years since publication 7.22; average citations per document 18.58; average citations per year per document 1.911. - Authorship and collaboration: 8998 unique authors, 13,330 author appearances; authors per document 2.61; co-authors per document 3.87; single-authored documents 336 (~9.74%); collaboration index 2.81. - Country productivity (overall 1995–2020, incl. conferences): Top producers include US (592), China (490), UK (253), Germany (222), Spain (135), Italy (130), Netherlands (113), Canada (111), Japan (107), India (106). China moved from absent (1995–2005) to 2nd (2006–2015) and became most productive in 2016–2020, surpassing the US. - Citations: Overall top cited countries: US (22,225; 35.58%), UK (6458; 10.34%), Germany (4454; 7.13%), China (3897; 6.24%), Netherlands (3231; 5.17%), Canada (3242; 5.19%), Australia (2721; 4.36%), Switzerland (1672; 2.68%), Singapore (1621; 2.60%), Italy (1598; 2.56%). China rose to 2nd most cited in 2016–2020. - International collaboration rates (multiple-country publications, MCP): UK 32.41%, Germany 31.98%, Netherlands 28.32%, Spain 22.22%, Italy 21.54%, China 20.82%, Canada 18.92%, Japan 15.89%, US 14.36%, India 8.49%. - Strongest bilateral ties among MCPs (share of each country’s MCPs involving the partner): UK–Germany very strong (e.g., 97.18% of Germany’s MCPs involve UK; 84.15% of UK’s involve Germany). US–China also strong (US: 63.53% of MCPs with China, 54/85; China: 52.94%, 54/102). US–UK next (US: 49.41%; China–UK: 51.22%). - Collaboration clusters: Three clusters with two major hubs: a US/Asia-Pacific cluster (US, China, Singapore, Japan, India, etc.) and a European cluster (UK, Germany, Netherlands, others), plus a European-dominant cluster including Italy and Spain. UK and US have highest centralities (UK betweenness 59.58; closeness 0.034; US betweenness 48.94; closeness 0.034). - Thematic evolution: 1995–2005 focused on mental health/affective disorders with neuroimaging (e.g., depression, bipolar, schizophrenia) alongside affective computing as a developing core theme. 2006–2015 saw the rise of sentiment analysis, NLP, opinion mining, facial expression analysis, semantic web, gamification. 2016–2020 shifted toward commercial/pragmatic applications and methods (feature extraction, AI, VR), fine-tuning algorithms (pattern recognition, feature selection, EEG, physiological signals), and increased computing power/data availability. Affective computing became a basic/core transversal theme; facial expression became a transversal topic. - Modalities and methods: Growth in affective BCI (emotion detection/stimulation; neuroprosthesis advances). Text-based affect uses hybrid symbolic + sub-symbolic (SenticNet; Hourglass and OCC models). Voice-based processing trends toward end-to-end learning (audEERING). Vision/biosignal work employs weakly supervised learning. Multimodal fusion and transfer learning gaining traction; MuSe challenge as a benchmark initiative. - Leading authors: Consistent influence of Rosalind Picard (MIT). 2006–2020 productivity led by Bjoern Schuller (audio/voice affect) and Erik Cambria (sentiment analysis). Post-2016 includes Fuji Ren (linguistic-based affect; Wi-Fi CSI behavior understanding). Local citation leaders shift from Picard (1995–2005) to Pantic, Schuller, Cambria, etc. - Top-cited works include Zeng et al. (2009) survey (1436 cites), DEAP database (2012; 1101), Picard et al. (2001; 1062), and others spanning multimodal recognition, neuroimaging, and content-based retrieval. - Geopolitics and coverage: China’s node in citation networks shows growing centrality and recent average citation years (early 2010s), indicating recent influence. Japan underrepresented among top-cited countries possibly due to language and structural factors. Russia notably absent among top producers/collaborators (only 36 WoS items), with reasons posited (sanctions, militarization, funding/bureaucracy, language, publication venues), though private-sector initiatives may increase future output.

The bibliometric mapping addresses the research question by quantifying growth, identifying major national actors and alliances, and revealing thematic shifts. Findings challenge the narrative of a strict US–China AI cold war by demonstrating strong co-authorship ties and interdependent clusters spanning the US/Asia-Pacific and Europe, with knowledge transfer from established to emerging economies. The field has moved from clinical detection of mental illnesses toward commercially oriented applications and methodological refinement, reflecting market and computational developments. Methodologically, multimodal fusion, hybrid symbolic–subsymbolic approaches, end-to-end learning, and transfer learning are converging, with benchmark initiatives (MuSe) galvanizing collaboration between academia and industry. Theoretical debates on emotion (Ekman vs. constructionist models) underscore the need for culturally aware datasets and models; text-based affective computing currently leads in operationalizing richer emotion theories (OCC, Hourglass, SenticNet), while vision and biosignal modalities still often rely on basic emotion taxonomies. The global analysis highlights China’s rapid ascent in productivity and recent citations, suggesting a multipolar research landscape. The observed collaborations and cross-cultural considerations are significant for developing fair, robust emotion-sensing systems and for guiding policy on data governance and ethical deployment.

Affective computing has experienced rapid, exponential growth (11.36% annually) over 25 years, evolving from clinically oriented themes to commercially driven applications and advanced methodologies, with a unifying trend toward multimodal and transfer learning approaches. The geopolitical landscape is shifting: while the US remains highly cited and central, China has become the most productive (2016–2020) and increasingly influential, with strong international collaborations, particularly with the US. Two dominant collaboration clusters—US/Asia-Pacific and Europe—anchor the field’s social structure. The study contributes a comprehensive, reproducible bibliometric map of outputs, collaborations, and thematic evolution, providing a baseline for future research and policy. Future directions include operationalizing constructionist emotion theories into computational models, developing culturally adaptive datasets and benchmarks, expanding multimodal fusion and transfer learning, advancing affective BCI, and conducting cross-cultural studies on emotion acculturation to improve generalizability and ethical robustness.

- Database and language bias: Only Web of Science–indexed, English-language records were included, potentially omitting non-English or regionally indexed work, which may underrepresent countries like Japan and Russia. - Keyword-based retrieval: Reliance on the term “affective computing” may exclude relevant emotional AI research published under different keywords. - Temporal cutoff: Early access 2021 materials were removed; citation counts favor older publications. - Coverage and document types: Conference proceedings are included but their impact differs from journals; disciplinary and venue differences may affect comparability. - Bibliometric measures: Co-authorship and citation networks indicate influence and collaboration but do not capture research quality or societal impact; centrality metrics are sensitive to dataset boundaries.