Discovering compositional, interpretable structure in social behavior

A decomposed linear dynamical system (dLDS) represents mouse social action as a signed, graded mixture of shared linear operators, capturing 21 keypoint-MoSeq syllables with just three operators in single-mouse data.

• Why it matters: Categorical action-at-a-time models limit expressive quantification of social behavior, which is combinatorial and compositional across multiple animals and time.
• What they did: The authors apply dLDS to single-mouse and two-mouse datasets, comparing it with keypoint-MoSeq, and in the resident–intruder assay they infer primitives without behavioral labels.
• The result: In two-mouse data, dLDS separates self-action, inter-animal-geometry, and social-action, and it decodes human-labeled behaviors (“attack”, “investigation”, “mount”, “other”) more accurately than behavior-selective keypoint-MoSeq using fewer units.

AI summary of the abstract · may contain errors · not medical advice · read the paper.