Figure 6. Stereotyped and distinct action sequences characterize stages of web-building.

A. HHMM design. Within each parent state (regime), transitions are modeled as simple Markov processes characterized by a movement motif transition matrix (child states). Transitions between regimes (A_X → B_Y, etc.) by design can only occur through parent nodes, (i.e. A_X → X → Y → B_Y) creating a hierarchical structure.

B. Likelihood for in-sample (left) or out-sample (right) HHMMs for different regime numbers. Gray vertical lines represent the standard deviation of model likelihoods.

C. F₁ scores for the top 5 models for each regime-number limit. The top-ranked 5-regime model is highlighted in red, and used for D-F. Model accuracy (F₁ score) defined as the agreement between manual stage annotations and stage boundaries predicted by each HHMM. F₁ scores were defined based on the optimal mapping of HHMM predicted regimes to manual regimes. For models with fewer than 5 regimes, some predicted regimes were assigned to multiple web stages. The F₁ score for the combined proto/radii stage was computed by first merging proto-web and radii annotations for both predicted and manual regimes.

D. Spider positions for all regime occupations and webs for the 5-regime HHMM highlighted in red in (C). Spider positions were assigned to a regime based on the most likely regime designation for that time-point. Dashed red circles indicate stabilimenta.

E. Web-stage manually predicted by spider position versus regime state (bottom panel) predicted by the 5-regime HHMM highlighted in red in (C) for a typical web progression. P(R) is the probability of each regime state. Centroid trajectories corresponding to the maximum-likelihood regime estimate are displayed (right panels), with colors matching the bottom panel.

F. Same as (E), but for an atypical web progression.

See also Figure S6.