Discovering Social and Aesthetic Categories of Avatars: A Bottom-Up Artificial Intelligence Approach Using Image Clustering

This study applies unsupervised image clustering to 191 player-created avatars from a custom RPG avatar creator, asking whether machine learning can uncover meaningful aesthetic and social categories without relying on predefined labels. Using two methods, non-negative matrix factorization (NMF) and archetypal analysis (AA), the authors examine avatar images represented in several color spaces, including alpha, brightness, and the AHSB representation chosen for its alignment with human color perception. The central argument is that this bottom-up, image-based approach can reveal both explicit categories such as gender and implicit player preferences, offering game developers quantifiable insight into player needs and design limitations.

The findings indicate that several model configurations produced statistically meaningful associations with avatar attributes. The NMF and AA models that performed best (p < .05) with respect to strength, endurance, and wisdom were AB-NMF-k6, A-NMF-k6, SB-AA-k3, and AHSB-AA-k6. Across these models, six clusters emerged as the optimal balance between interpretability and correlation strength. Some clusters contained only male or only female avatars, and visual choices such as dark clothing were linked to particular ability allocations, for example male avatars associated with high physical but low magic attributes.

The analysis reflects both role and gender stereotypes in player behavior, with male avatars tending toward physical strength and lower wisdom. The authors interpret these patterns as evidence that unsupervised clustering can capture design and identity structures that surface in player choices. The scope is confined to a single RPG avatar creator and a sample of 191 avatars, so the generalizability of these categories beyond that context remains an open question.

The study positions AI clustering as a bottom-up tool for modeling player values and identity. Future directions include text-based topic modeling, demographic filtering, and applications to procedural game design, which would extend the method beyond visual analysis.