Figure 10. Recall-normalised confusion matrices for leave-one-subject-out (LOSO) evaluation of the pooled linear classifier across all seven participants. Rows represent true emotion labels and columns predicted labels. Values show recall percentages; colour intensity ranges from white (0%) to dark red (100%). Each panel corresponds to one held-out participant and is annotated with overall LOSO accuracy.

The pooled classifier displays participant-specific absorption patterns, in which multiple emotions from the held-out participant are systematically reassigned to a single pooled category learned from other participants. Examples include P6, where pooled Contempt absorbs Happiness, Neutral, Sadness, and portions of Anger and Disgust; P7, where pooled Flow absorbs Fear and Anger; P1, where pooled Anger absorbs Neutral; and P3, where pooled Neutral absorbs Flow. These absorption patterns reveal structural collisions between participant-specific emotional geometries rather than gradual confusion between neighbouring categories.

Overall LOSO accuracy ranges from 16.7% (P3) to 33.2% (P1), with a mean of 26.91% across all seven participants. The heterogeneity of the absorption patterns suggests that pooled failure is not attributable to a common classification boundary or model limitation. Instead, different participants' emotional manifolds occupy incompatible regions of feature space, producing participant-specific category remappings when geometrically distinct manifolds are forced into a shared representational framework. Hover over cells for exact values and absorption annotations.