Figure 1 — Analytical Pipeline
Analytical pipeline diagram Analytical pipeline showing temporal modelling, manifold analysis, frame-level analysis, cross-participant audit, and convergent evidence integration. 76-dimensional ARKit input 60fps · 7 participants · 9 emotions Continuous temporal facial capture Temporal modelling BiLSTM+Attention 60-frame motion windows Manifold analysis UMAP geometry Participant-specific + audit Frame-level analysis LDA projection Linear classification Personalised ≈99% Pooled ≈27% Large generalisation gap Centroids Trajectories + velocity Shared embedding audit PCA clouds + phase map LDA dim-1: 65–100% Linear personalised: 96.35% Pooled LOSO: 26.91% Convergent evidence geometry · dynamics · classification phenomenology · cross-participant audit Individual manifolds Section 4 Cross-participant geometry Section 5 Dynamical organisation Sections 6–7
Figure 1. Analytical pipeline for the study. Continuous 76-dimensional ARKit facial-motion data were recorded at 60fps across seven participants expressing nine emotion categories. Three primary analytical streams were applied to the same dataset: supervised temporal modelling using a BiLSTM+Attention classifier, participant-specific manifold analysis using UMAP trajectory geometry and velocity-field methods, and supervised frame-level analysis using Linear Discriminant Analysis (LDA) and linear classification. Findings were then subjected to cross-participant audit through shared-embedding reconstruction, PCA-normalised trajectory comparison, and leave-one-subject-out classification. Convergent evidence across geometry, dynamics, classification, phenomenology, and audit supports the central claim of the paper: emotion categories remain identifiable within individuals, but their geometric organisation is participant-specific.