August 16, 2026
IEEE ICDM 2026 (Research Track)
Bayesian Spectral Emotion Transition Discovery from Multi-Annotator Disagreement
Keito Inoshita, Takato Ueno
A two-stage unsupervised framework (BSETD) that treats annotator disagreement as signal rather than noise, discovering emotion-transition structure directly from multi-annotator soft labels. Bayesian transition estimation (a hierarchical Dirichlet-Multinomial model with FDR control) and graph-spectral decomposition separate emotional inertia from contagion, validated across five corpora and cross-lingual data and aligned with established findings in psychology.